**Photoacoustic Tomography Appearance of Fat Necrosis: A First-in-Human Demonstration of Biochemical Signatures along with Histological Correlation**

**Yonggeng Goh 1,†, Ghayathri Balasundaram 2,†, Hui Min Tan 3, Thomas Choudary Putti 3, Celene Wei Qi Ng 4, Eric Fang 1, Renzhe Bi 2, Siau Wei Tang 4, Shaik Ahmad Buhari 4, Mikael Hartman 4, Ching Wan Chan 4, Yi Ting Lim 1, Malini Olivo 2,\* and Swee Tian Quek 1,\***


**Abstract:** A 50-year-old woman with no past medical history presented with a left anterior chest wall mass that was clinically soft, mobile, and non-tender. A targeted ultrasound (US) showed findings suggestive of a lipoma. However, focal "mass-like" nodules seen within the inferior portion suggested malignant transformation of a lipomatous lesion called for cross sectional imaging, such as MRI or invasive biopsy or excision for histological confirmation. A T1-weighted image demonstrated a large lipoma that has a central fat-containing region surrounded by an irregular hypointense rim in the inferior portion, confirming the benignity of the lipoma. An ultrasound-guided photoacoustic imaging (PA) of the excised specimen to derive the biochemical distribution demonstrated the "masslike" hypoechoic regions on US as fat-containing, suggestive of benignity of lesion, rather than fat-replacing suggestive of malignancy. The case showed the potential of PA as an adjunct to US in improving the diagnostic confidence in lesion characterization.

**Keywords:** photoacoustic; lipoma; ultrasound; MRI

Fat necrosis, or cell death of adipose tissue, is a common benign condition that occurs from the lack of oxygen supply to adipose tissue [1]. As common causes include trauma or post-surgical changes [2], fat necrosis often presents as a palpable soft tissue mass at superficial regions [3].

Ultrasound (US) is the first-line imaging tool for these superficial lesions, but imaging appearances are extremely varied [4] due to the age of the lesion, which manifests as varying degrees of hardening, fibrosis, and degeneration. This often results in a diagnostic dilemma, which necessitates further cross-sectional imaging or invasive procedures, such as biopsy or excision for histological confirmation (Figures 1 and 2). There is, hence, an unmet clinical need for an adjunct imaging modality to US to improve diagnostic capability.

**Citation:** Goh, Y.; Balasundaram, G.; Tan, H.M.; Putti, T.C.; Ng, C.W.Q.; Fang, E.; Bi, R.; Tang, S.W.; Buhari, S.A.; Hartman, M.; et al. Photoacoustic Tomography Appearance of Fat Necrosis: A First-in-Human Demonstration of Biochemical Signatures along with Histological Correlation. *Diagnostics* **2022**, *12*, 2456. https://doi.org/ 10.3390/diagnostics12102456

Academic Editor: Viktor Dremin

Received: 20 September 2022 Accepted: 10 October 2022 Published: 11 October 2022

**Publisher's Note:** MDPI stays neutral with regard to jurisdictional claims in published maps and institutional affiliations.

**Copyright:** © 2022 by the authors. Licensee MDPI, Basel, Switzerland. This article is an open access article distributed under the terms and conditions of the Creative Commons Attribution (CC BY) license (https:// creativecommons.org/licenses/by/ 4.0/).

**Figure 1.** A 50-year-old woman with no past medical history presented with a left anterior chest wall mass. Clinically, the mass was soft, mobile, and non-tender. (**A**): US image of the left anterior chest wall mass shows a wider than tall mass with well-circumscribed margins (bold arrows). It contains internal fat echogenicity and is in keeping with a fat-containing lesion (i.e., lipoma). Within the inferior portions (dotted arow), there are focal "mass-like" hypoechoic nodules seen, which are surrounded by a hyperechoic capsule. (**B**): Doppler US of the focal "mass-like" nodules performed, demonstrated mild increased peripheral vascularity around the hyperechoic capsule but no increased vascularity within the hypoechoic "mass-like" nodules. Findings were indeterminate for malignant change of a lipomatous lesion.

**Figure 2. MRI of left anterior chest wall.** MRI was performed for the patient in view of possible malignant change (**A**): Axial T1w image of the left anterior chest wall mass shows a fat-containing lesion in the left anterior chest wall (bold arrows) in keeping with a lipoma. Within the inferior portion, there are fat-containing areas on MRI, which are surrounded by T1w hypointense irregular bands (dotted arrows). (**B**): These irregular T1w hypointense bands show enhancement on the post contrast enhanced image. (**C**): Sagittal post contrast enhanced image shows similar findings of irregular rim enhancement around a focal fat-containing central component within inferior portions of the lipoma.

Photoacoustic (PA) tomography, a hybrid optical imaging modality, is based on the light-induced ultrasound waves providing the contrast of optical imaging combined with the high spatial resolution of ultrasound [5]. Its ability to provide the distribution of endogenous chromophores, such as blood oxygenation [6], water [7], lipid [7–10] and recently collagen [11,12], makes it attractive as a potential adjunct tool to various aspects of ultrasound. This is particularly useful in regions abundant with these chromophores, such as superficial soft tissues and the breast, or in fat-containing and fibrotic/necrotic conditions, such as fat necrosis. However, these theoretical advantages and usages of chromophore differentiations have not been demonstrated in daily clinical usage. Herein, we present interesting images that demonstrate for the first time the biochemical signatures of fat necrosis derived by PA and its agreement with histopathology (Figure 3). A detailed description of PA imaging protocol and image reconstruction is included in Appendix A. This work showcases the potential of PA as an adjunct for US to improve the diagnostic confidence for fat necrosis.

**Figure 3.** PA imaging of the excised tissue. Patient underwent uneventful excision of the left chest wall mass and was imaged with PA. Methods are as described in Appendix A. (**A**): Gross pathology of a representative cut section of the lipomatous tumor showing yellowish fatty appearance and foci of necrosis (dotted arrows). Focal cystic change containing oily fluid secondary to fat necrosis was present at where the "mass-like" nodules were seen on US (bold arrows). This area is rimmed by fibrosis. (**B**): H&E-stained microscopic image of the lipomatous tumor confirms prominent degenerative change, featuring areas of fat necrosis (dotted arrows), cystic change (bold arrows), and fibrosis. (**C**): shows zoomed in image of the cyst with septation (diamond arrow). This largest area of cystic change (bold arrows) corresponds to the "mass-like" nodule as seen on ultrasound (labelled as \*). (**D**): The "mass-like" nodule near the posterior margin (\*), surrounded by a hyperechoic rim (bold arrow), was targeted for PA imaging. Corresponding PA images showing distribution of lipid (**E**) and collagen (**F**), or their overlay (**G**) demonstrated collagenous signal corresponding to the hyperechoic halo in keeping with fibrosis. Within the "mass-like" region, the imaged portions demonstrated lipid signal, which was similar in intensity to the surrounding lipoma. No suspicious fat or collagenous-replacing masses were identified. No blood signals were obtained from this ex vivo study as there was no active ongoing blood flow after lesion excision. Scale bar 5 mm.

In this case, the authors have successfully demonstrated biochemical features of fat necrosis on PA as a first in-human demonstration with histopathological correlation. PA was able to identify the "mass-like" hypoechoic regions on US as fat-containing, rather than fat-replacing. On pathology, the lipid signals on PA correspond to liquefied necrotic fat from cystic degeneration, while the collagen signals on PA correspond to the fibrosis around the cavity. Hence, the biochemical capability of fat and collagen characterization could help to resolve ambiguous findings on US and improve diagnostic confidence for fat necrosis.

As it is crucial to obtain pathological correlation with US-PA images, excised tissues with no active blood signals had to be obtained. The authors believe the incorporation of blood signals in in vivo imaging would further intensify our understanding of the pathophysiology and respective imaging correlations for fat necrosis. Although the results from a single case may seem promising, more work must be done to validate these findings. In particular, more work must be done to validate findings of benign lipomas, fat necrosis, and malignant lipomatous tumors to investigate their biochemical differences. With more data, PA could potentially translate downstream into clinical imaging workflows for better characterization of superficial/breast lumps where fat-containing lesions are common. However, for widespread clinical adoption, there needs to be vast improvements in both hardware and software for PA imaging to improve its imaging depth (to at least 3–4 cm), its field of view, as well as spectral coloring.

**Author Contributions:** Conceptualization, Y.G.; methodology, Y.G., G.B., H.M.T., C.W.Q.N., E.F., R.B., M.H., S.W.T., C.W.C., M.H., T.C.P., S.A.B. and Y.T.L.; software, Y.G., G.B. and H.M.T.; writing—original draft preparation, Y.G.; writing—review and editing, Y.G., G.B. and H.M.T.; supervision, M.O. and S.T.Q. All authors have read and agreed to the published version of the manuscript.

**Funding:** This research was funded by (1) Agency of Science, Technology and Research (A\*STAR), under its BMRC Central Research Fund (UIBR) 2021 and (2) NMRC clinician-scientist individual research grant new investigator grant (CS-IRG NIG).

**Institutional Review Board Statement:** The study was conducted in accordance with the Declaration of Helsinki and approved by the Institutional Review Board (or Ethics Committee) of NATIONAL UNIVERSITY HOSPITAL, SINGAPORE (2017/00805, date of approval: 12 October 2017) for studies involving humans.

**Informed Consent Statement:** Informed consent was obtained from all subjects involved in the study.

**Conflicts of Interest:** The authors declare no conflict of interest. The funders had no role in the design of the study; in the collection, analyses, or interpretation of data; in the writing of the manuscript, or in the decision to publish the results.

#### **Appendix A**

#### *Appendix A.1. Materials and Methods*

The excised lipoma was collected fresh from the operating theatre and subjected to US-PA imaging using the commercially available MSOT inVision 512-echo system (iThera Medical GmbH, Munich, Germany), fitted to a customized 2D handheld probe (specifications: 2D array of 256 detector elements (arranged along a 125◦ arc on a spherical surface: radius 40 mm and 5 MHz centre frequency)). The detector was placed on the inferior portions of the lipoma to image the "mass-like" nodules.

PA images were acquired at near infrared wavelengths—700, 730, 760, 800, 850, 920, 930, 970, 1000, 1050, 1064, and 1100 nm—to allow for deep-tissue imaging. Big differences in absorption of light at these wavelengths by chromophores in breast tissue, such as blood, lipid, and collagen, have aided the unmixing of these chromophores. Data were processed using ViewMSOT 3.8 software (Release 3.8, Munich, Germany) and reconstructed using the backprojection algorithm, after applying a bandpass filter with cut-off frequencies of

50 kHz and 6.5 MHz. The distribution of collagen and lipid were visualized through spectral unmixing of the reconstructed data.

#### **References**


## *Article* **Combination of Optical Biopsy with Patient Data for Improvement of Skin Tumor Identification**

**Yulia Khristoforova 1,\*, Ivan Bratchenko 1, Lyudmila Bratchenko 1, Alexander Moryatov 2, Sergey Kozlov 2, Oleg Kaganov <sup>2</sup> and Valery Zakharov <sup>1</sup>**


**Abstract:** In this study, patient data were combined with Raman and autofluorescence spectral parameters for more accurate identification of skin tumors. The spectral and patient data of skin tumors were classified by projection on latent structures and discriminant analysis. The importance of patient risk factors was determined using statistical improvement of ROC AUCs when spectral parameters were combined with risk factors. Gender, age and tumor localization were found significant for classification of malignant versus benign neoplasms, resulting in improvement of ROC AUCs from 0.610 to 0.818 (*p* < 0.05). To distinguish melanoma versus pigmented skin tumors, the same factors significantly improved ROC AUCs from 0.709 to 0.810 (*p* < 0.05) when analyzed together according to the spectral data, but insignificantly (*p* > 0.05) when analyzed individually. For classification of melanoma versus seborrheic keratosis, no statistical improvement of ROC AUC was observed when the patient data were added to the spectral data. In all three classification models, additional risk factors such as occupational hazards, family history, sun exposure, size, and personal history did not statistically improve the ROC AUCs. In summary, combined analysis of spectral and patient data can be significant for certain diagnostic tasks: patient data demonstrated the distribution of skin tumor incidence in different demographic groups, whereas tumors within each group were distinguished using the spectral differences.

**Keywords:** Raman spectroscopy; cancer risk factors; skin cancer; PLS analysis; statistical significance

#### **1. Introduction**

The annually growing trend of melanoma disease is observed worldwide [1]. Research [2] estimated that 106,110 new cases of melanoma were diagnosed and about 7180 people died of this disease in the USA in 2021. The growth of melanoma can be caused by different personal [3–5], behavioral, and socioeconomic factors [6,7]. The National Cancer Institute has reported [2,8–10] that melanoma is more common in men than women and more frequent among whites in comparison with other races or ethnicities. Moreover, there is a strong relationship between melanoma cases and patient age [2]. For example, incidence rates for MM skin cancer in the UK are the highest in people aged 75 and over [11].

In terms of environmental factors, ultraviolet radiation is the most dangerous factor causing melanoma growth [11,12]. Localization can also be a potentially informative factor for more accurate skin cancer diagnosis, because some types of skin tumor often develop in the body areas that are directly exposed to UV radiation, with others appearing in covered body sites subjected to intense sunburn because of their rare exposure to regular UV radiation [13].

High risk can also be associated with family history: about 10% patients with melanoma have a family history of the disease [14–17]. The study by Hemminki et al. [18] demonstrated that melanoma is several times more common in people whose first-degree relatives

**Citation:** Khristoforova, Y.; Bratchenko, I.; Bratchenko, L.; Moryatov, A.; Kozlov, S.; Kaganov, O.; Zakharov, V. Combination of Optical Biopsy with Patient Data for Improvement of Skin Tumor Identification. *Diagnostics* **2022**, *12*, 2503. https://doi.org/10.3390/ diagnostics12102503

Academic Editor: Viktor Dremin

Received: 7 September 2022 Accepted: 12 October 2022 Published: 15 October 2022

**Publisher's Note:** MDPI stays neutral with regard to jurisdictional claims in published maps and institutional affiliations.

**Copyright:** © 2022 by the authors. Licensee MDPI, Basel, Switzerland. This article is an open access article distributed under the terms and conditions of the Creative Commons Attribution (CC BY) license (https:// creativecommons.org/licenses/by/ 4.0/).

have had melanoma. Moreover, researchers [19,20] reported on the relationship between patient history of skin neoplasm and the risk of developing melanoma, suggesting that the patient's history indicates risk of skin cancer growth. People working in certain professions can have a higher risk of skin cancer and some precancer conditions, due to interaction with dangerous industrial carcinogens [21].

Preliminary diagnosis of melanoma using dermoscopy [22] or other developing optical biopsy techniques [23–25] did not consider the above risk factors that may be potential prerequisites for developing skin cancer. However, incorporating patient-specific information can improve the accuracy of disease identification based on clinical studies [26–28]. Pacheco and Krohling [26] demonstrated the importance of clinical features for skin cancer detection based on clinical images and confirmed the hypothesis that patient clinical information is important for this task. However, they concluded that the clinical features they examined were not practical indicators for all types of skin lesions. Zeng et al. [27] examined skin tumors using Raman spectral data, considering various risk factors, and revealed that only patient age significantly contributed to improved diagnosis of malignant tumors. Taking into account the findings of other research teams [26–28], we aimed to test the possibility of improving skin cancer identification with our experimental data, by combining Raman and autofluorescence data as well as patient information.

In our previous work [29,30], we performed an optical biopsy using Raman and autofluorescence (AF) spectroscopy to diagnose skin cancer. Raman spectroscopy has proved to be a sensitive research instrument in clinical practice for a number of purposes [24,27,30,31]. The proposed method [29,32] was able to classify skin neoplasms with a mean accuracy higher than the accuracy of general practitioners or trainees, and with comparable or less accuracy than trained dermatologists and experts. Therefore, it remains necessary to improve the accuracy of skin cancer diagnosis performed with Raman and AF analysis.

The aim of our study was to estimate the prognostic possibility of combining individual patient factors with the results of optical biopsy for detecting skin cancer. The spectral data of 617 skin tumors that were analyzed in our previous work [29] were combined with data on risk factors, for joint analysis. We demonstrated the results of the proposed approach by combining Raman spectroscopy and AF with individual patient factors such as environmental risk, history, and personal risk factors for classification of malignant skin tumors, melanoma, and other skin neoplasms.

#### **2. Materials and Methods**

#### *2.1. Experimental Setup*

The detailed description of the experimental setup for simultaneous Raman and AF signal registration was presented in our previous studies [29,32]. The scattering spectral response from skin tissue in the near-infrared region was stimulated using a thermally stabilized diode laser module (LuxxMaster, LML-785.0RB-04, PD-LD, Ushio Inc., Tokyo, Japan) with 785 ± 0.1 nm central wavelength. The laser power density on the skin was about 0.3 W/cm<sup>2</sup> and did not cause any damage to skin or discomfort in patients. The optical Raman probe (RPB785, InPhotonics Inc., Norwood, MA, USA) contained supplying and collecting branches. Laser excitation at 785 nm was delivered to the skin surface by means of the excitation optic fiber (0.22 NA, 100 μm) and the supplying branch of the probe with a band-pass filter and a focusing lens. The scattered radiation was collected by the same lens and delivered to the collecting branch by the dichroic mirror and the conventional mirror. The longpass filter cut the excitation laser wavelength from the collected signal, and the Raman and fluorescence signals of skin tissue were transmitted to the spectrometer using the focusing lens and the collecting fiber (NA 0.22, 200 μm). The collected signal was decomposed into a spectrum using a portable spectrometer (QE65Pro, Ocean Optics Inc., Largo, FL, USA). The spectra were registered in the 780–1000 nm region with spectral resolution of 0.2 nm. The acquisition time was 20 s with a triple accumulation. The QE65Pro detector was cooled down to −15 ◦C. The silicon tip on the probe provided the 7–8 mm distance between the skin surface and the probe for all measurements.

#### *2.2. Patients*

The protocols of the in vivo tissue diagnostics were approved by the ethical committee of Samara State Medical University (Samara Region, Samara, Russia, protocol No 132, 29 May 2013), the clinical studies fall within The Code of Ethics of a Doctor of Russia, approved at the 4th Conference of the Russian Medical Association, and within the World Medical Association Declaration of Helsinki. The study involved 615 patients of different ages, including 178 men and 437 women, who consulted specialized oncologists in the Samara Regional Clinical Oncology Dispensary from May 2017 to December 2019. All the patients were aged ≥ 18. Informed consent was acquired from all patients before the in vivo study.

Spectral measurements of 617 tumors were carried out for 615 patients. The spectral measurement of each skin tumor was registered from the approximate central point of the tumor area. The region of interest for spectral registration of tumors was confirmed by a medical specialist on the basis of dermatoscopic images. The skin tumors were localized at different body sites. The sizes of skin tumors varied widely, from 0.3 to 5 cm. Summary of the patients and tumors is presented in [29]. In accordance with results of histopathological analysis, the analyzed spectral cohort included 204 malignant tumors (70 malignant melanomas (MM), 122 basal cell carcinomas (BCC) and 12 squamous cell carcinomas (SCC)), as well as 413 benign tumors (26 dermatofibromas (DF), 62 papillomas (PP), 40 hemangiomas (HE), 113 seborrhoeic keratosis (SK), 170 nevi (NE) (all types), 1 cutaneous horn, and 1 benign tumor of epidermal appendage).

#### *2.3. Risk Factors for Skin Cancer Growth*

Cancer develops when human cells are damaged due to various factors and the number of damaged cells starts to grow uncontrollably. In this work, we analyzed several risk factors that can potentially provoke skin cancer growth.

At the initial appointment, the oncologist collected the patient history and potential risk factors for skin cancer growth: gender (G), age (A), tumor localization (L), family history (FH), personal history (PH), sun exposure (SE), size (S), and occupational hazards (OH). All the collected demographic indicators were defined by the patient survey. However, for different reasons, not every patient provided the full set of collected risk factors. Only gender, age and localization factors were received for all the 617 skin neoplasm spectra. Therefore, we considered two spectra datasets:


All the risk factors were digitized:


The digitization of the patient factors was performed by specialized oncologists at the Samara Regional Clinical Oncology Dispensary.

#### *2.4. Preprocessing and Statistical Analysis of Spectra*

The spectra were recorded in the 780–1000 nm region, but only the 803–914 nm spectral region corresponding to the 300–1800 cm−<sup>1</sup> wavenumber region in terms of Raman spectroscopy was analyzed. Firstly, the raw spectra in the region of interest (803–914 nm) were preprocessed by the following process: smoothing by the Savitsky–Golay filter, normalization by the standard normal variate method (SNV), and centering.

In accordance with the data described in Section 2.2, we considered six classification models with different sets of risk factors:


Each spectrum included the Raman and AF signals in the region of interest (of 803–914 nm) and, therefore, represented a discrete set of intensity values at the 515 wavelengths (in accordance with the spectral resolution of the spectrometer). For the subsequent regression analysis, the 515 spectral parameters respectively representing each tumor after preprocessing were combined with the corresponding risk factor parameters. Therefore, in classification models (I.1), (I.2), and (I.3) each tumor was represented as 518 predictors (515 spectral parameters and three risk factor parameters) for PLS analysis, and in classification models (II.1), (II.2), and (II.3) as 523 predictors (515 spectral parameters and eight risk factor parameters), respectively.

The experimental data were processed using partial least square discriminant analysis (PLS-DA) [33]. The PLS-DA method was applied to build a regression model between the analyzed tumor predictors and tumor types. Stability of the PLS-DA classification was checked by means of 10-fold cross-validation. The number of latent variables (LVs) for the PLS-DA models was chosen according to the minimum of the RMSE in the 10-fold cross-validation. To estimate the importance of all tumor predictors in the model, variable importance in projection (VIP) analysis was performed [34]. The VIP scores highlighted the informative predictors of tumors in the regression model that were more important for classifying different tumor types. Higher relative intensity of VIP score indicated that the predicted variable was more significant. To determine the differentiation accuracy of the tumor analysis, the PLS predictors were calculated as numeric values of tumor diagnosis in the built regression model.

The results of the skin tumor differentiation were visualized using a bee-swarm diagram and the receiver operating characteristic (ROC) curves plotted using R studio software [35]. The ROC analysis shows the diagnostic performances of the regression model. For quantitative analysis, the area under the curve (AUC) was calculated. The significance of the AUCs and the comparisons between different AUCs were tested in a standard manner [36].

#### **3. Results**

#### *3.1. Malignant vs. Benign Neoplasms*

(I.1) To discriminate the malignant (n = 204) vs. benign (n = 413) neoplasms from set (I), the 0.600 (0.567–0.652) ROC AUC was obtained using only the spectral data (RS and AF data). The complementation of spectral dataset (I) with three risk factors made it possible to improve the ROC AUCs to 0.818 (0.778–0.841). Moreover, adding each patient factor separately to the spectral data significantly increased the ROC AUC (see Table 1). The distribution of VIP scores as a weighted sum of loadings is shown in Figure 1, highlighting all spectral features for all loadings obtained in this PLS classification model. For this model, the VIP scores were utilized to classify malignant versus benign tumors by determination of informative predictors (gender (G), age (A), location (L), and 515 spectral parameters) in

regression specification. The VIP scores presented in Figure 1 demonstrate that age (A) is the most informative risk factor, which was proved by the most significant improvement of ROC AUC (0.804, *<sup>p</sup>* = 9 × <sup>10</sup>−9) when only age was incorporated into the spectral data, compared with the other factors.

**Table 1.** Results of regression models.


**Figure 1.** VIP scores for PLS-DA model: classification of malignant (n = 204) vs. benign (n = 413) neoplasms (I.1); importance values of risk factors and spectral factors are plotted along different horizontal axes because of a wide value scatter. (G) gender; (A) age; (L) localization.

The ROC AUCs and the bee-swarm diagram for this classification are presented in Table 1 and Figure 2a–c.

(II.1) For set (II), the classification of malignant (n = 157) vs. benign (n = 324) neoplasms using the PLS analysis was performed with the 0.610 (0.556–0.663) ROC AUC on the basis of only the spectral data, and with the 0.789 (0.746–0.832) ROC AUC when supplying the spectral data with eight risk factors. For this set, age was also the most important risk factor. The ROC AUCs and bee-swarm diagram are presented in Table 1 and Figure 2d–f.

Table 1 presents the ROC AUCs of the models built using all risk factors separately. Improvement of the ROC AUC by incorporating the spectral data with all risk factors to identify malignant skin cancer was statistically significant (*p* < 0.05) in models I.1 and II.2.

#### *3.2. MM vs. Benign Pigmented Neoplasms (Ne and SK)*

(I.2) In this classification task, regression analysis of the cases from dataset (I) using only the spectral data was performed with 0.690 (0.630–0.761) ROC AUC. For this task, the combined analysis of the spectral data and the three risk factors significantly improved the diagnostic performance to 0.825 (0.766–0.884) ROC AUC. The contribution of all three risk factors in this model was significant for MM identification (*p* < 0.05), whereas separately adding age, gender, or localization did not result in significant improvement of the ROC AUC. Figure 2g–i and Table 1 show the results from cohort (I) for this classification task.

(II.2) In the same classification task for cohort (II), MMs (n = 49) were differentiated from benign pigmented neoplasms (n = 221) with 0.789 (0.718–0.861) ROC AUC using only the Raman and AF spectral data, and 0.849 (0.785–0.914) ROC AUC when combining the spectral and risk factor variables. However, in this case, there were no significant differences (*p* = 0.14) between the ROC AUCs obtained for the models with the eight risk factors and those without. Figure 2 presents only the statistically significant results and therefore does not include a diagram for this model. Table 1 indicates the ROC AUCs for this classification task.

**Figure 2.** Results of diagnostic models with statistical significance (*p* < 0.05). I.1: Malignant vs. benign neoplasms: (**a**) ROC AUCs, bee-swarm diagrams of PLS predictors for tumor classification based on (**b**) spectral parameters and (**c**) combination of spectral parameters and three patient factors. II.1: Malignant vs. benign neoplasms: (**d**) ROC AUCs, bee-swarm diagrams of PLS predictors for tumor classification based on (**e**) spectral parameters and (**f**) combination of spectral parameters and eight patient factors. I.2: MM vs. benign pigmented neoplasms: (**g**) ROC AUCs, bee-swarm diagrams of PLS predictors for tumor classification based on (**h**) spectral parameters and (**i**) combination of spectral parameters and three patient factors.

#### *3.3. MM vs. SK*

(I.3) When classifying the MMs (n = 70) vs. SKs (n = 113) from cohort (I) using the spectral data, 0.791 (0.722–0.859) ROC AUC was obtained. When the three risk factors were combined with the Raman and AF spectral data, the ROC AUC increased to 0.844 (0.786–0.902) but with no statistical significance.

(II.3) When analyzing dataset (II) with the same classification task, the discrimination model showed 0.820 (0.748–0.892) ROC AUC obtained by combining spectral data and the eight risk factors, while the analysis of only the spectral data was performed with 0.814 (0.740–0.888) ROC AUC.

The AUCs of this classification task were insignificantly improved (*p* > 0.05) in models I.3 and II.3. The diagnostic performance results for this classification task are presented in Table 1.

#### **4. Discussion**

The classification results for different types of skin neoplasms based on Raman and AF spectral data were demonstrated in our previous research [29,32]. In this current work, we combined individual patient risk factors and spectral data to obtain a more precise skin cancer diagnosis, in particular of malignant tumor and MM. It should be noted that true statistics of skin cancer incidence might differ from the data we obtained, due to including in this study only those patients who were aware of skin tumors, attentive to their health, and had access to resources for tumor detection.

Considering the significance of the analyzed patient factors, our classification models have demonstrated that although gender was a significant factor for classifying malignant versus benign skin tumors in both model I.1 and model II.1, it was not significant for diagnosis of MM. The analyzed cohort was heterogeneous in the numbers of men and women: women outnumbered men 2–3 times in the general cohort and in the analyzed classes. However, the proportion of men with malignant tumors among all men involved in this study was 0.43, whereas the relative proportion of women was 0.28. At the same time, the relative incidence rates of MM were 0.13 among men and 0.11 among women (Figure S1a). The statistics for different ethnicities and races vary. According to statistics from Australia and the USA [10,13], the incidence of MM is higher among men than women. In Russia in 2020, the standardized incidence rates (number per 100,000 population) of skin cancer (without MM) and MM among men were 21.48 and 4.08 respectively, whereas among women the figures were 20.62 (skin cancer without MM) and 4.32 (MM) [37]. Data for our cohort was collected from May 2017 to December 2019, and revealed that in our study the relative number of malignant cases was higher among men, while the number of MM cases was the same among men and women.

Localization was also a significant factor in models I.1 and II.1 for malignant versus benign classification. We suppose that the significance of this factor can be explained by the most common localization group for different tumor types (Figure S1b). In model I.1, most of the malignant tumors, namely 86 out of 204 cases (about 42%), were located on the head and neck, while 209 out of 413 benign tumors (about 51%) were situated on the trunk. Despite the fact that more MM cases occurred on the trunk (about 51%), the large number of malignant tumors on the head and the neck was due to the contribution of the BCC and SCC cases. However, when classifying MM and benign pigmented tumors or SK, this factor was found to be insignificant, because most cases within each class occurred on the trunk: 51% among melanoma cases, 51% among benign pigmented cases, and 47% among cases of seborrheic keratosis.

The sun exposure factor was insignificant in all models (II.1, II.2, and II.3), but exposure to sun radiation is partly responsible for localization. BCC and SCC are more likely to occur on body areas that are most exposed to solar radiation, i.e., on the head. According to our data presented in Figure S1, 61% of BCC and 71% of SCC in this study were localized on the head and the neck [29]. On the other hand, most MM (51%) and other melanocytic tumors, such as pigmented nevus, occurred on the trunk and legs: these body areas may be subjected to intense sunburn because of less frequent exposure to regular UV radiation. Other research [38] reports that trunk melanomas are more strongly associated with pigmented nevus counts. Thus, exposure to sun radiation is an equally important growth factor for all melanocytic tumors, confirmed by a similar distribution of various melanocytic tumors (e.g., most cases of MM and benign pigmented nevus were recorded on the trunk). Therefore, these tumors were not distinguished within each localization.

Our models suggest that age is a significant factor when classifyin malignant and benign tumors, because the patient distribution in each age group was different. Most cases of malignant tumors (about 44%) were recorded in patients over 70, while the groups of patients with benign tumors aged from 30 to 39, from 40 to 49, from 50 to 59, and from 60 to 69 were equal in size. For classification of MM versus only SK, age was completely insignificant and did not improve the ROC AUC when added to the spectral data, because the distribution of patients in age groups for MM and SK was fairly similar (Figure S1c). The maximum frequency of MM and benign pigmented tumors (Ne and SK) was recorded in the age group from 60 to 69. The larger number of pigmented benign tumors in the age group from 60 to 69 was due to more SK cases, which did not allow significant separation of these classes by age. However, there were differences in the numbers of patients in other age groups: among those with benign pigmented tumors, more than 30% were under 40 because of a larger number of young patients with pigmented nevus. This resulted in the fact that adding the age factor to the spectral data improved the ROC AUC to 0.808 (0.734–0.881) for the classification of MM versus benign pigmented tumors, but these differences were not sufficient for statistical significance.

Finally, it should be noted that in all three classification tasks (models II.1, II.2, and II.3), OH, FH, and PH were not able to improve the ROC AUCs (see Table 1). We lack precise data on the behavioral and genetic factors obtained in the survey because the oncologists collected this information by questioning the patients, which can lead to inaccuracy and uncertainty. To enhance the importance of these factors, they could be defined in a more precise way, for example, in terms of controlling exposure to sunlight or to chemicals that can be dangerous in case of contact with skin. Thereby, the significance of such patient data as gender, age, tumor localization, and size can become more reliable for tumor detection.

Diagnostic performance combining patients' demographic data with optical data has been evaluated in several works [26–28]. Zhao et al. [27] investigated whether incorporating such patient demographics as gender, skin type, localization, and age into Raman spectral analysis can improve performance in malignant skin cancer diagnosis. Using PLS analysis, the authors reported that the ROC AUC improved significantly from 0.913 (0.892–0.933) to 0.934 (0.917–0.952) (*p* < 0.05) after combining only the Raman data with all the demographics, to differentiate malignant and benign skin lesions. In comparison with the study by Zeng et al., in our work we analyzed a larger set of risk factors for cancer growth, including not only demographics but patient lifestyle and behavioral factors as well. We similarly found that combining all the risk factors with the spectral data achieved better performance in discriminating malignant and benign tumors, increasing the ROC AUC from 0.600 to 0.818 with three factors and from 0.610 to 0.789 with eight factors. We increased the ROC AUCs by 30–36% taking into account the patient risk factors, while in the study by Zeng et al. [27] the improvement was only 2%. Probably, this greater improvement of malignant skin cancer identification was a result of a different signal-to-noise ratio in the spectral data. Our spectral data were recorded with a lower signal-to-noise ratio [39] that resulted in low accuracy of detecting malignant neoplasms by only the Raman and AF spectra (0.600 and 0.610 ROC AUC in models I.1 and II.1, respectively), and a significant improvement when the patient factors were added. In the previous work [27], the authors used a highly sensitive spectroscopic system that allowed them to obtain a high ROC AUC 0.913 using only the Raman spectral data.

Kharazmi et al. [28] proposed a non-invasive fast BCC detection tool that incorporates dermoscopic lesion features and clinical patient information including lesion localization, size, and elevation, as well as patient age and gender. The integrated analysis of the patient profile and dermoscopic features using data-driven feature learning allowed them to increase the ROC AUC for BCC detection from 0.847 to 0.911, in comparison with only the dermoscopic features. According to our statistics [29], BCC cases occur within specific demographic conditions, for example, 61% of BCCs were located on the head and neck, and 90% of BCCs were recorded from patients over 60. Thus, it might be assumed that we would be able to determine the significance of patient factors for BCC detection. However, we analyzed BCC and MM as malignant tumors and did not estimate the importance of patient information for identifying BCC only. Considering that our statistical results about BCC are in good agreement with the statistics reported by Kharazmi et al., this may suggest the significance of patient factors only for several types of skin lesions.

It is also interesting to compare the results of the proposed methodology and the results of dermoscopic image analysis performed by dermatologists, which represents the current standard for clinical diagnosis of skin lesions. According to research [40], the accuracy of melanoma vs. non-melanoma skin lesion classification was 79.9% for novice dermatologists, 83.3% for qualified dermatologists, and 86.9% for experts. The mean diagnostics performance of 21 board-certified dermatologists using dermoscopic images to classify 71 malignant vs. 40 benign lesions was nearly 71% sensitivity and 81% specificity [41]. Thus, the proposed methodology can classify skin neoplasms with a mean accuracy higher than GPs and trainees, but with slightly less accuracy than trained dermatologists and experts.

To sum up, our results show that information on patient risk factors and Raman and AF spectral data can complement each other to provide more accurate skin cancer identification. For each skin tumor type, we observed a specific distribution trend by gender, age group, and localization, in good agreement with worldwide statistics on skin tumor incidence. Patients' age and tumor localization are able to discriminate tumors in different groups, but these factors become insignificant when analyzing different skin tumors within individual groups. For example, similar numbers of malignant, benign, pigmented tumors, SK, and MM were recorded on the trunk or in patients aged from 50 to 59 or from 60 to 69. So, within a separate demographic group, accuracy results for different tumor type diagnosis can differ when only the Raman and AF spectral data are used. For this reason, to differentiate malignant versus benign skin tumors we improved the ROC AUCs by adding risk factors to the model. To differentiate MM versus pigmented skin tumors or SK, similar demographic trends did not allow us to increase the performance accuracy of skin tumor identification. To improve diagnostic performance, the proposed methodology may be added to the estimation of neoplasm morphology performed during dermoscopy analysis. Deep learning-based applications using computer visualization have shown promising results in detecting melanoma based on the analysis of dermoscopic images [40–42]. However, additional studies are required to estimate the capability of joint dermoscopy analysis and low-cost Raman systems.

#### **5. Conclusions**

We tested the possibility of improving skin cancer detection by combining spectral analysis with analysis of individual patient characteristics and factors for skin cancer growth. We analyzed two cohorts of patients with skin tumors: (I) the cohort with 617 spectra of different tumors and three patient factors for each case, and (II) the cohort with 481 spectra of different tumors and eight risk factors. For each cohort, three classification tasks were considered: malignant versus benign tumors, MM versus benign pigmented tumors, and MM versus SK.

The significance of risk factors for type of cancer growth was estimated when all factors were combined with the spectral data, and when each factor was added separately to the Raman and AF spectral data. Statistical improvement was achieved for the classification of 204 malignant tumors and 413 benign tumors, from 0.610 to 0.818 ROC AUC, *<sup>p</sup>* = 2 × <sup>10</sup><sup>−</sup>11, when spectral data in the 300–1800 cm−<sup>1</sup> range were combined with three individual patient factors for skin cancer growth. Moreover, classification of 157 malignant tumors and 324 benign tumors using the spectral data and eight risk factors was statistically improved from 0.610 to 0.789, *<sup>p</sup>* = 5 × <sup>10</sup><sup>−</sup>7. Finally, 70 MMs and 283 benign pigmented skin neoplasms were differentiated with a statistical improvement from 0.709 to 0.825, *p* = 0.02 when combining the spectral data and the three risk factors. Improvements of ROC AUC for discriminating MM (n = 49) and pigmented benign tumors (n = 172) with eight factors, MM (n = 70) and SK

(n = 113) with three factors, and the MM (n = 49) and SK (n = 90) with eight factors were all statistically insignificant.

Our results show that among all risk factors, patient demographics including gender, age, and tumor localization were statistically significant for detecting skin tumor type, due to their univocal definition. In contrast, the data for behavioral factors were collected by staff directly from patients and might therefore lack accuracy. For certain classification tasks, it was found that the combination of spectral data and patient risk factors was significant. Particular overall trends for each skin tumor type were observed for patient age, gender, and tumor localization. However, these demographic features did not allow us to discriminate different tumor types, especially pigmented tumors, within an individual demographic group. Therefore, distinguishing skin tumors in groups with similar demographics was possible using the Raman and AF spectral data only. However, these findings need to be verified in further experimental cohort studies.

**Supplementary Materials:** The following supporting information can be downloaded at: https: //www.mdpi.com/article/10.3390/diagnostics12102503/s1, Figure S1: Patients' statistics: (**a**) distribution by gender; (**b**) distribution by localization; (**c**) distribution by age groups.

**Author Contributions:** Conceptualization, I.B., V.Z.; Methodology, Y.K., I.B., L.B.; Validation, I.B., V.Z.; Investigation, Y.K., L.B., A.M.; Resources, S.K., V.Z.; Data Curation, I.B., A.M.; Writing—Original Draft Preparation, Y.K.; Writing—Review & Editing, I.B., V.Z.; Supervision, V.Z., S.K., O.K.; Project Administration, V.Z., O.K.; Funding Acquisition, I.B. All authors have read and agreed to the published version of the manuscript.

**Funding:** This work was supported by the Russian Science Foundation under Grant No. 21-75-10097, https://rscf.ru/project/21-75-10097/.

**Institutional Review Board Statement:** The protocols of the in vivo tissue diagnostics were approved by the ethical committee of Samara State Medical University (Samara Region, Samara, Russia, protocol No 132, 29 May 2013); clinical studies fall within The Code of Ethics of a Doctor of Russia, approved at the 4th conference of the Russian Medical Association, and within the World Medical Association Declaration of Helsinki.

**Informed Consent Statement:** Informed consent was obtained from all subjects involved in the study.

**Data Availability Statement:** The data presented in this study are available on request from the corresponding author.

**Conflicts of Interest:** The authors declare no competing interests.

#### **References**


p. 252. ISBN 978-5-85502-268-1. Available online: https://oncology-association.ru/wp-content/uploads/2021/11/zis-2020 -elektronnaya-versiya.pdf. (In Russian)


## *Article* **Study of the Natural Crystalline Lens Characteristics Using Dual-Energy Computed Tomography**

**Jeffrey R. Sachs 1, Javier A. Nahmias 2, Kevin D. Hiatt 1, James G. Bomar 1, Thomas G. West 1, Paul M. Bunch 1, Marc D. Benayoun 1, Chris Lack <sup>1</sup> and Atalie C. Thompson 2,\***


**Abstract:** There is a paucity of radiologic literature regarding age-related cataract, and little is known about any differences in the imaging appearance of the natural crystalline lens on computed tomography (CT) exams among different demographic groups. In this retrospective review of 198 eyes in 103 adults who underwent dual-energy computed tomography (DECT) exams of the head, regions of interest spanning 3–5 mm were placed over the center of the lens, and the x-ray attenuation of each lens was recorded in Hounsfield Units (HU) at 3 energy levels: 40 keV, 70 keV, and 190 keV. Generalized estimating equations (GEEs) were used to assess the association of clinical or demographic data with lens attenuation. The mean HU values were significantly lower for the older vs. younger group at 40 keV (GEE *p*-value = 0.022), but there was no significant difference at higher energy levels (*p* > 0.05). Mean HU values were significantly higher for females vs. males and non-whites vs. non-Hispanic whites at all 3 energy levels in bivariate and multivariable analyses (all *p*-value < 0.05). There was no significant association between lens attenuation and either diabetes or smoking status. The crystalline lens of females and non-whites had higher attenuation on DECT which may suggest higher density or increased concentration of materials like calcium and increased potential for cataract formation. Given the large scope of cataracts as a cause of visual impairment and the racial disparities that exist in its detection and treatment, further investigation into the role of opportunistic imaging to detect cataract formation is warranted.

**Keywords:** dual-energy computed tomography; natural crystalline lens; sex; race

#### **1. Introduction**

Cataracts are the leading cause of blindness worldwide and affect over 24 million Americans [1]. Of the subtypes of age-related cataracts, nuclear sclerotic cataracts are the most common [2], and are thought to result from cumulative lifetime exposure to a range of insults to the ocular lens including ultraviolet light, ocular trauma, ocular surgery, corticosteroid use, radiation exposure, smoking, and diabetes mellitus.

Currently, cataracts are almost universally a clinical diagnosis, made by confirming opacification or discoloration of the lens during slit lamp examination by a trained specialist in optometry or ophthalmology. However, barriers to access of basic healthcare services have caused many patients to seek acute medical care through the emergency room, which provides an opportunity for incidental discovery of comorbid chronic conditions, such as cataracts [3].

Computed tomography (CT) scans are one of the most common medical tests performed in the emergency room [4]. CT has proven useful in diagnosing traumatic cataracts [5–8], wherein the lens becomes hypoattenuating due to increased fluid content. However, there is a striking paucity of radiologic literature on the imaging characteristics of

**Citation:** Sachs, J.R.; Nahmias, J.A.; Hiatt, K.D.; Bomar, J.G.; West, T.G.; Bunch, P.M.; Benayoun, M.D.; Lack, C.; Thompson, A.C. Study of the Natural Crystalline Lens Characteristics Using Dual-Energy Computed Tomography. *Diagnostics* **2022**, *12*, 2857. https://doi.org/ 10.3390/diagnostics12112857

Academic Editor: Viktor Dremin

Received: 22 September 2022 Accepted: 15 November 2022 Published: 18 November 2022

**Publisher's Note:** MDPI stays neutral with regard to jurisdictional claims in published maps and institutional affiliations.

**Copyright:** © 2022 by the authors. Licensee MDPI, Basel, Switzerland. This article is an open access article distributed under the terms and conditions of the Creative Commons Attribution (CC BY) license (https:// creativecommons.org/licenses/by/ 4.0/).

25

sage-related cataracts, which are far more common than traumatic cataracts [2]. Moreover, little is known about differences in the characteristics of the natural lens on CT imaging among different demographic groups.

One reason for this lack of knowledge is that when using conventional single-energy CT techniques, the lenses are bland, relatively homogeneous structures that garner little attention unless they are displaced, dysmorphic, or absent. Moreover, since the natural lens is a radiosensitive organ at particular risk for radiation effects due to its superficial location in the human body, intentional imaging of the lens during radiologic image acquisition is not generally recommended [9]. In clinical practice, however, the lens is often incidentally included in the CT scan field of view. This circumstance presents the radiologist with an opportunity to assess the lens for pathology such as cataracts.

In recent years, dual-energy CT (DECT) has risen to the forefront of CT imaging acquisition due to its ability to capitalize on the differences in energy-dependent x-ray absorption of different materials within the patient by using low- and high-energy x-ray spectra. DECT techniques have numerous recently described applications in the practice of neuroradiology [10]. Low kilo–electron volt (keV) (e.g., 40 keV) virtual monoenergetic imaging (VMI) techniques allow for improved contrast-to-noise ratio among soft tissues of similar attenuation, even in the absence of iodinated contrast media [11]. High keV (e.g., 190 keV) VMI have been used to describe unique attenuation patterns of silicone oil which can be found in the eye after certain retinal surgeries [12,13].

Given the ability of DECT to improve discrimination between a material of high atomic number (such as calcium or iodine) and a material of low atomic number (such as hemorrhage) [14], we hypothesized that it would be useful in detecting differences in lens composition that could be related to early cataract formation. The lens is made of proteins which are low in atomic number. If the cataractic lens were to calcify, then one would expect that the attenuation of the lens would increase fairly dramatically over time, especially at 40 keV, due to the combined impact of low energy incident photons and high atomic number on the photoelectric effect, and subsequently on x-ray beam attenuation. Although total lens calcification is rare [15], lens calcium content has been shown to correlate with the degree of opacification of cataractous lenses [16]. Impaired intracellular Ca+2 signaling in lens epithelial cells is known to play an intrinsic role in both cortical and nuclear cataractogenesis [17]. Moreover, we might expect nuclear cataracts to have an increased density due to the pathologic aggregation and compaction of the proteins of the nuclear fibers which has been observed on histology [2]. However, to our knowledge, no study has investigated the characteristics of the natural crystalline lens on DECT or whether there are differences in attenuation of the lens in different demographic groups or at different levels of energy.

In this retrospective study, we examined whether there is a relationship between possible demographic and clinical risk factors for cataract formation, such as age, race, sex, diabetes and smoking status, with the X-ray attenuation of the natural crystalline lens on DECT of the head at 40, 70, and 190 keV VMI. Such imaging analyses of the crystalline lens may provide critical foundational understanding of the changes in the lens that one might observe on DECT as well as help identify groups at potential risk for cataract formation.

#### **2. Materials and Methods**

#### *2.1. Subjects*

For this retrospective, Health Insurance Portability and Accountability Act-compliant, institutional review board-approved study, a local institutional radiology database (syngo.via, Siemens Healthineers) was queried with the goal of identifying eligible lenses for study inclusion. Screening eligibility criteria were (1) age ≥18 years and (2) received non-contrast DECT of the head at our institution between July and December 2020. A younger adult (age ≥18 and <30 years) and older adult cohort (age >70 years) were collected to facilitate assessment of any age-related differences in the lens characteristics. Patients were excluded if there was evidence of acute orbital trauma (including periorbital hematoma, orbital

hemorrhage, and orbital fractures). Pseudophakic eyes were not included in the analysis. Using these criteria, a total of 198 eyes in 103 adult subjects were identified. Demographic or clinical risk factors for cataract such as age, sex, race/ethnicity, smoking status, and history of diabetes were collected from the electronic medical record. Documentation of known cataract status was also collected when available.

#### *2.2. Image Acquisition*

All non-contrast DECT examinations of the head were performed utilizing a dualsource system (SOMATOM Drive or Flash; Siemens Healthineers, Erlangen, Germany). The DECT acquisition parameters were as follows: 80 kVp/Sn140 kVp acquisitions, Quality Reference mAs of 400/200, pitch of 0.70, rotation time of 0.5 s, and with automated tube current modulation, CareDose 4D.

#### *2.3. Image Analysis*

All dual-energy post-processing and region of interest (ROI) analysis was performed in syngo.via (Siemens Healthineers) using the Monoenergetic+ application. Measurements were obtained across all subjects using an image slice thickness of 1 mm. Two-dimensional circular ROIs were drawn by a trained radiology resident or neuroradiology fellow in the center of the native lenses. Care was taken to identify the center of the lens and avoid any streak or beam hardening artifacts. An attending neuroradiologist with 5 years subspecialty experience reviewed and optimized the ROI position as needed. All ROIs were sized to be between 3–5 mm2. A representative example of how ROI measurements were obtained is given in Figure 1. After the initial ROI was placed, the ROI location did not change while HU attenuation measurements were recorded at each of the 40 keV, 70 keV, and 190 keV energy levels. The mean attenuation and standard deviation (SD) were recorded in Hounsfield Units (HU) for each ROI at each energy level.

**Figure 1.** A representative example of how ROI measurements were obtained for the study. All measurements were obtained in the axial plane using 1 mm thick slices. ROI data was obtained using the default syngo.via Monoenergetic+ application. All scans were non-contrast dual-source, dual-energy CT exams. See the text Section 2.2. for the DECT acquisition parameters.

#### *2.4. Statistical Analyses*

Descriptive statistics were used to describe the clinical and demographic characteristics of the population. Pearson's correlation was used to compare the attenuation in Hounsfield units between right and left eyes among subjects with both natural lenses present on DECT. In order to account for the correlation of both eyes within a given subject, separate generalized estimating equations (GEEs) were constructed to assess the bivariate relationship between each of the demographic or clinical variables and the relative attenuation of the lens at each energy level (40, 70, and 190 keV). Factors that were significant in bivariate analyses (*p*-value < 0.05) were entered into multivariable GEE models of the HU for each energy level. All analyses were performed with Stata (version 17.0, StataCorp, College Station, TX, USA). A *p*-value < 0.05 was considered statistically significant.

#### **3. Results**

#### *3.1. Subjects*

Included subjects were sub-categorized into an older adult (N = 53; mean age 81.4 ± 5.7 years; range 73–101) and younger adult cohort (N = 50; mean age 22.66 ± 2.93 years; range 18–27). Approximately 46.6% of subjects were female, 38.8% self-identified as non-Hispanic white, 22.3% had diabetes, and 37.9% were never smokers (Table 1). Among the older adults, only 19 subjects (29 eyes) had ophthalmic examination data of the natural lens in their electronic healthcare record so that cataract status could be determined.

**Table 1.** Clinical and demographic characteristics of 103 adult subjects who underwent Dual Energy Computed Tomography (DE-CT).


#### *3.2. Image Analysis*

Ninety-five subjects had a natural crystalline lens present in both eyes. There was a strong and significant positive correlation between the HU attenuation values in the right and left eye at each of the three energy levels (all *p* < 0.0001) (Table 2).

**Table 2.** Correlation of right and left eye attenuation measurements on Dual-Energy Computed Tomography in 95 pairs of eyes in 95 subjects.


DECT = Dual-Energy Computed Tomography. Bolded *p*-values are statistically significant (*p* < 0.05).

Table 3 displays the bivariate association between clinical and demographic characteristics and attenuation of the natural crystalline lens in the full cohort (198 eyes in 103 subjects). At 40 keV, the older adults had a significantly lower average HU than younger adults (70.88 ± 13.63 vs. 75.70 ± 13.21; *p* = 0.020), but there was no significant difference in the measurements for older vs. younger adults at the 70 keV or 190 keV energy levels (both *p* > 0.05). However, in the subgroup of older adults with clinical documentation of a cataract (N = 19 subjects with 29 eyes), the Hounsfield unit values were higher, measuring 74.65 ± 14.57) at 40 keV, 79.05 ± 7.58 at 70 keV, and 80.85 ± 9.29 at 190 keV.


**Table 3.** Bivariate analysis of clinical and demographic characteristics associated with natural crystalline lens attenuation measured at three energy levels on Dual Energy Computed Tomography (DE-CT) (N = 198 eyes in 103 subjects).

Generalized estimating equations = GEE. Bolded *p*-values are statistically significant (*p* < 0.05).

At all three energy levels, the mean HU attenuation values were significantly higher for females vs. males (40 keV: 76.4 ± 13.3 vs. 70.7 ± 13.4; 70 keV: 79.0 ± 7.4 vs. 74.2 ± 7.6; 190 keV: 80 ± 7.9 vs. 75.8 ± 9; all *p* <= 0.01) and for non-whites vs. non-Hispanic whites (40 keV: 77.6 ± 13.1 vs. 70.5 ± 13.2; 70 keV: 80 ± 7.9 vs. 74.1 ± 7.0; 190 keV: 81.0 ± 8.1 vs. 75.5 ± 8.5; all *p* <= 0.001). No significant association was detected between lens attenuation and either diabetes or smoking status (all *p* > 0.05).

In multivariable analyses, both female sex (all *p* < 0.01) and nonwhite race/ethnicity (all *p* < 0.01) but not age-group (*p* > 0.05) remained significant independent predictors of lens attenuation at all 3 energy levels (Table 4).

**Table 4.** Multivariable generalized estimating equations of characteristics associated with natural crystalline lens attenuation measured at three energy levels on Dual Energy Computed Tomography (DE-CT) (N = 198 eyes in 103 subjects).


GEE = Generalized estimating equations; CI = confidence interval. Bolded *p*-values are statistically significant (*p* < 0.05).

#### **4. Discussion**

To our knowledge, this study is the first to demonstrate the ability of DECT to detect differences in the attenuation of the crystalline lens on DECT among demographic and clinical groups that may be at increased risk for cataract formation.

We found that the crystalline lens of females and non-white subjects had significantly higher attenuation on DECT at all energy levels. This finding may suggest a higher density lens or increased concentration of materials like calcium which could be related to lens opacification from early cataract formation [16]. The Salisbury Eye Evaluation Study found that African Americans had higher rates of cortical opacity and progression of cortical cataracts [18]. The Age-Related Eye Disease study similarly found that both females and nonwhites were at greater risk of cortical cataract formation [19]. One hypothesis in females is that age-related withdrawal of estrogen may play a role in the progression cataracts through the loss of its anti-oxidative effects [20]. Women also have a higher prevalence of osteoporosis which has been associated with increased risk for cataract, presumably through common pathways of impaired calcium homeostasis [21]. Such impaired calcium signaling in lens epithelial cells can result in increased cytosolic calcium concentration [22] which predisposes to cortical cataracts in particular. For example, one study demonstrated that the total calcium in lenses with cortical cataracts measured four times higher than in clear lenses [23]. Cortical cataracts can also have discrete calcium deposits which would have substantially higher attenuation on DECT. Moreover, impaired calcium signaling can also lead to opacification of the nuclear portion of the lens resulting in a mixed cataract type [17]. Interestingly, the subset of older adults with a documented cataract also had higher HU values. However, the retrospective design of the study limited more direct assessment of the relationship between these characteristics and cataract formation since ophthalmic examination data was not available in a majority of patients. Nevertheless, the study demonstrated that DECT has the potential to identify differences in lens attenuation among groups at risk for cataract. Such findings could form the rationale for a future prospective study in which patients undergoing DECT of the head are recruited for ophthalmology examination to determine whether a clinically and visually significant opacification of the lens is present.

To our surprise, no significant association between lens attenuation and either diabetes or smoking status was detected (all *p* > 0.05). High blood glucose in the setting of poorly controlled diabetes can lead to generation of polyols that result in increased osmotic stress in the lens fibers causing them to swell and rupture [24]. Given previous studies on traumatic cataracts and the low HU attenuation from increased water concentration, one might have expected diabetic patients to have a lower HU attenuation. The lack of an association here could be related to the fact that severity of diabetes was not able to be determined since HbA1c was not consistently available. Future studies could collect additional lab criteria to stratify by diabetic control. Similarly, prior studies have associated smoking with an increased risk of nuclear sclerotic type cataracts due to increased oxidative stress [25,26]. It is possible that smoking data from the medical record is less accurate than questionnaire data which could be collected in a future prospective study.

The older adult cohort in general trended toward a lower attenuation of the lens at 40 keV, but this relationship was not significant in multivariable analyses or at higher energy levels. It is possible that DECT did not identify a Hounsfield-unit based threshold to distinguish lens age because different subtypes of cataract may result in different alterations in lens attenuation. For example, some cataracts might decrease lens attenuation if there is an increase in fluid whereas others may increase lens attenuation if there is an increase in deposition of calcium or other high-density materials. Future studies could consider whether specific subtypes of cataract have different degrees of lens attenuation on DECT.

Although an ROI based methodology to confirm cataract by DECT is not yet possible, the lens remains a potentially attractive target for artificial intelligence-based segmentation given its simple shape, near-uniform size, and the high level of contrast between the lens itself and the surrounding fluid of the aqueous and vitreous humor [27]. Larger DECT imaging datasets pooled across multiple institutions may make possible the training

and development of deep learning algorithms for predicting cataract risk. Additionally, while obtaining the study data we noticed occasional patients that demonstrated highly variable attenuation values in the lenses that produced a "speckled" pattern of attenuation (Figure 2). Such a pattern suggests the usefulness of texture analysis for identification of cataracts, where differences in spatial heterogeneity within the ROI of a cataract lens can be distinguished from a non-cataract lens, despite having, for example, similar overall mean HU attenuation values. In fact, prior studies have reported greater than 90% accuracy in identifying cataracts from eye photographs when using lens ROI uniformity (a texture analysis metric) as a training feature for a nearest-neighbor classifier [28], though to the best of our knowledge this has not yet been reproduced with DECT images.

**Figure 2.** Axial non-contrast DECT image zoomed for lens detail. An 84-year-old lens demonstrates a "speckled" appearance—with scattered punctate foci of increased attenuation. This pattern of spatial heterogeneity raises the question of whether texture analysis may be a more useful method of assessing for cataract at CT.

A majority of patients in our study did not have historic ophthalmic data, suggesting there could be gaps in access to ophthalmic care that could be potentially addressed during an emergency department visit. Racial disparities in access to cataract surgery are welldocumented, especially among black patients [29]. Inequities in the proportion of patients that carry health insurance affects access to routine health care and utilization [30,31]. Racial minorities are also known to be less likely than white patients to have a primary care provider and are more likely to rely on the Emergency Department for routine care needs [32–34]. As such, radiologists could have an opportunity to help improve healthcare disparities through invention of novel opportunistic imaging screening techniques (i.e., screening an organ for pathology when that organ is incidentally imaged as part of a study obtained for a separate indication). While traditional cataract evaluation will remain the gold-standard for cataract diagnosis, future studies should investigate whether patients presenting to the ED who undergo DE-CT could benefit from opportunistic detection of cataract and development of a care pathway that refers such patients to ophthalmology for confirmation and treatment.

Our study has several limitations. In order to allow for examination of possible agerelated differences in the lens, two cohorts of older and younger adults were collected. Thus, the study findings are not generalizable to middle-aged patients. Achieving accurate lens cortex attenuation values is problematic due to the concern for volume averaging with adjacent fluid filled structures. However, if eyes had cortical changes overlying the nucleus these would have been detected within the ROI. Additional ophthalmic clinical data as to

the presence or grade of cataract were only available in a subset of older adults. Based on the findings in this study, a future prospective study could be designed in which patients are referred to an ophthalmologist for clinical exam of the lens following incidental imaging of the crystalline lens on DECT during an emergency room encounter.

Another potential limitation of the study is the presence of beam hardening artifacts at 40 keV VMI. In our study, the authors took great care to avoid beam hardening artifacts emanating from the bony orbit; if such artifacts had been included in the ROI analysis, they would have led to spurious increases or decreases in HU values. Photon counting CT holds promise to both improve spectral resolution and reduce beam hardening artifacts in the future [35].

Our study was performed on one of two distinct types of DECT scanners, both dualsource systems manufactured by Siemens, and with nearly identical protocols. These results may not necessarily be generalizable to all commercially available forms of spectral CT imaging, though the underlying physics should remain the same. Moreover, several factors can affect the accuracy of the HU values, including the aforementioned beam hardening artifacts, spectral energy, convolution kernel, and patient positioning [36]. CT scanners are calibrated such that the HU value for pure water does not deviate more than 2 HU from the reference value of 0 [37], but the possibility exists that small differences in scanner calibration in conjunction with a combination of the above factors could impact the measurements taken.

#### **5. Conclusions**

Our study found that non-white and female patients had significantly higher lens attenuation values on DECT, which may suggest increased lens density, the presence of high-density materials like calcium, or a predisposition for cataracts. Future collection of ophthalmic examination data in patients with DECT findings will provide useful diagnostic confirmation of the clinical relevance of these findings. Given the large scope of cataracts as a cause of visual impairment and the racial disparities that exist in its detection and treatment, further investigation into the role of opportunistic DECT imaging to detect cataract formation is warranted. As these technologies continue to advance, incorporation of photon counting CT and artificial intelligence-based textural analysis may improve the utility of such radiologic studies in the clinical care of patients with incidental cataracts.

**Author Contributions:** Conceptualization, J.R.S., M.D.B., T.G.W., P.M.B., C.L., A.C.T.; Methodology, J.R.S., A.C.T., P.M.B., C.L., M.D.B., K.D.H., J.G.B., T.G.W.; Software, J.R.S., A.C.T.; Validation, J.R.S., A.C.T.; Formal analysis, J.R.S., A.C.T.; Investigation, J.R.S., K.D.H., J.G.B., A.C.T., J.A.N.; Resources, J.R.S., A.C.T.; Data curation, J.R.S., A.C.T., J.A.N.; Writing—original draft preparation, J.R.S., A.C.T., J.A.N.; Writing—review and editing, J.R.S., A.C.T., J.A.N., K.D.H., T.G.W., P.M.B., M.D.B., C.L., J.G.B.; Visualization, J.R.S.; Supervision, J.R.S., A.C.T.; Project administration, not applicable; Funding acquisition, not applicable. All authors have read and agreed to the published version of the manuscript.

**Funding:** This research received no external funding. Dr. Thompson receives support from the NIH/NEI (K23-EY030897) and the Wake Forest Older Americans Independence Center (P30- AG021332). Dr. Bunch received support as an AUR GERRAF fellow.

**Institutional Review Board Statement:** The study was conducted in accordance with the Declaration of Helsinki, and approved by the Institutional Review Board Wake Forest Baptist Medical Center (protocol code IRB00068983 and approved 10/5/2020).

**Informed Consent Statement:** A waiver of informed consent was granted by the Wake Forest Baptist Health Medical Center Institutional Review Board due to the retrospective nature of this study. All collected data were de-identified and study results were published in aggregate.

**Data Availability Statement:** Restrictions apply to the availability of these data. Data were obtained from the electronic healthcare record at Atrium-Wake Forest Baptist Health and a de-identified dataset may be available with the permission of Atrium-Wake Forest Baptist Health's legal team.

**Conflicts of Interest:** The authors declare no conflict of interest.

#### **References**

