Next Article in Journal
Multi-Source Image Fusion Based Regional Classification Method for Apple Diseases and Pests
Previous Article in Journal
Analysis of the Model Used to Predict Continuous Deformations of the Land Surface in Areas Subject to Discontinuous Deformations—A Case Study
Previous Article in Special Issue
Towards Metric-Driven Difference Detection between Receptive and Nonreceptive Endometrial Samples Using Automatic Histology Image Analysis
 
 
Search for Articles:
Title / Keyword
Author / Affiliation / Email
Journal
Article Type
 
 
Advanced Search
 
Section
Special Issue
Volume
Issue
Number
Page
 
Journals
Applied Sciences
Volume 14
Issue 17
10.3390/app14177696
applsci-logo
Submit to this Journal Review for this Journal Propose a Special Issue
► Article Menu

Article Menu

share Share announcement Help format_quote Cite thumb_up
...
Endorse textsms
...
Comment
first_page
settings
Order Article Reprints
Font Type:
Arial Georgia Verdana
Font Size:
Aa Aa Aa
Line Spacing:
Column Width:
Background:
Article

The Quantification of Myocardial Fibrosis on Human Histopathology Images by a Semi-Automatic Algorithm

by
Diana Gonciar
1,2,
Alexandru-George Berciu
3,*,
Alex Ede Danku
3,
Noemi Lorenzovici
3,
Eva-Henrietta Dulf
3,*,
Teodora Mocan
4,5,
Sorina-Melinda Nicula
6 and
Lucia Agoston-Coldea
1
1
2nd Department of Internal Medicine, Iuliu Hațieganu University of Medicine and Pharmacy, 400012 Cluj-Napoca, Romania
2
Department of Pathology, Iuliu Hațieganu University of Medicine and Pharmacy, 400012 Cluj-Napoca, Romania
3
Automation Department, Faculty of Automation and Computer Science, Energy Transition Research Center, Technical University of Cluj-Napoca, 400114 Cluj-Napoca, Romania
4
Physiology Department, Iuliu Hațieganu University of Medicine and Pharmacy, 400012 Cluj-Napoca, Romania
5
Department of Nanomedicine, Regional Institute of Gastroenterology and Hepatology, 400158 Cluj-Napoca, Romania
6
Department of Pathology, Săcele Municipal Hospital, 505600 Săcele, Romania
*
Authors to whom correspondence should be addressed.
Appl. Sci. 2024, 14(17), 7696; https://doi.org/10.3390/app14177696 (registering DOI)
Submission received: 7 August 2024 / Revised: 26 August 2024 / Accepted: 28 August 2024 / Published: 31 August 2024
(This article belongs to the Special Issue Recent Progress and Challenges of Artificial Intelligence in Bioinformatics and New Medicine)
Browse Figures
Versions Notes

Abstract

:
(1) Background: Considering the increasing workload of pathologists, computer-assisted methods have the potential to come to their aid. Considering the prognostic role of myocardial fibrosis, its precise quantification is essential. Currently, the evaluation is performed semi-quantitatively by the pathologist, a method exposed to the issues of subjectivity. The present research proposes validating a semi-automatic algorithm that aims to quantify myocardial fibrosis on microscopic images. (2) Methods: Forty digital images were selected from the slide collection of The Iowa Virtual Slidebox, from which the collagen volume fraction (CVF) was calculated using two semi-automatic methods: CIELAB-MATLAB® and CIELAB-Python. These involve the use of color difference analysis, using Delta E, in a rectangular region for CIELAB-Python and a region with a random geometric shape, determined by the user’s cursor movement, for CIELAB-MATLAB®. The comparison was made between the stereological evaluation and ImageJ. (3) Results: A total of 36 images were included in the study (n = 36), demonstrating a high, statistically significant correlation between stereology and ImageJ on the one hand, and the proposed methods on the other (p < 0.001). The mean CVF determined by the two methods shows a mean bias of 1.5% compared with stereology and 0.9% compared with ImageJ. Conclusions: The combined algorithm has a superior performance compared to the proposed methods, considered individually. Despite the relatively small mean bias, the limits of agreement are quite wide, reflecting the variability of the images included in the study.

1. Introduction

Heart failure is a condition with a strong negative impact on the survival of patients, as well as having a high financial burden due to its increased incidence, with a lifetime risk of occurrence of 20% [1]. The mechanism behind it is the cardiac remodeling process, which includes cardiomyocyte injury (hypertrophy, apoptosis and necrosis) and fibrosis [2]. In myocardial fibrosis, type I collagen predominantly accumulates (to the detriment of type III collagen), which is more stable and rigid; the consequences of this are increased myocardial stiffness, decreased contractility, diastolic dysfunction, decreased myocardial perfusion, and increased risk of arrhythmia [3]. Considering the prognostic role of fibrosis, its identification is essential [4]. Current assessment options include serum markers, magnetic resonance imaging, and endomyocardial biopsy [5]. The histological evaluation of fibrosis is still the gold standard, with the significant disadvantage that it may miss a focal process due to spatial heterogeneity and the fact that extensive sampling is not feasible [5].
In both endomyocardial biopsies and autopsy specimens, the assessment of fibrosis is part of the pathologist’s routine activities. Myocardial fibrosis is of two types: diffuse and replacement [6]. Replacement fibrosis involves the necrosis of cardiomyocytes and their replacement with scar tissue, thus leading to thick collagen fibers that form confluent areas of more than 2 mm, uninterrupted by the interposition of cardiomyocytes [7]. Diffuse fibrosis includes interstitial and perivascular fibrosis, resulting in the accumulation of fine collagen fibers (of approximately 10–500 µm) surrounding groups of cardiomyocytes or individual cells, without forming confluent areas [7]. Diffuse fibrosis comprises three patterns, which most often coexist: microscars (small areas of replacement fibrosis arising in the context of isolated cardiomyocyte death), perivascular deposits (around arterioles and intramyocardial vascular branches), and deposits around the perimysium and endomysium [8].
The extent of fibrosis can be quantified by measuring the collagen volume fraction (CVF), which represents the ratio of the proportion of collagen and the totality of the examined myocardial tissue, being reported to have a prognostic role in patients with hypertrophic cardiomyopathy [9]. The assessment can also be performed semi-quantitatively, although the threshold values for defining severe fibrosis are highly variable from one study to another [5], without a consensus at the moment. Although special stains for collagen fibers are used to evaluate fibrosis, the hematoxylin–eosin (H&E) stain is routinely performed in laboratories, being inexpensive and rapid. It allows fibrosis visualization, although it provides low contrast compared to other stains [10].
As a result of digitization, numerous computational methods have been developed to assist physicians and, at the same time, to encourage standardization [11,12,13,14]. Several methods to quantify fibrosis on microscopic images have been reported, but most are focused on the liver [15,16] and kidney [17]. Additionally, the technical details available in the literature are few, and the validation is frequently performed via semi-quantitative comparison performed by the pathologist, which is influenced by subjectivity and brings an additional volume of work [18]. As a result of the significant development of omic technologies, the quantitative information generated based on the analyzed biological processes requires the integration and modeling of the data, an area where machine learning strategies can have an impact [19,20]. In cardiovascular pathology, a computational sequencing method was applied for the spatial transcriptomics of the aorta dissection, succeeding in the spatial localization of different cell subpopulations [21]. Employing a technique based on graph contrastive learning and multi-task learning can enhance spatial localization in transcriptomics [22]. Single-cell RNA sequencing is the preferred method for highlighting distinct, unknown cells and might also be performed to evaluate intercellular communication and the mechanisms of fibrogenesis in the heart [23]. Still, the performance of studies in this field is limited by the missing values arising from biological considerations [24]. Machine learning methods [25] or deep learning models are proposed to tackle this challenge [26].
The present paper proposes the validation of a machine learning algorithm developed on digital images, which aims to quantify collagen deposition in the myocardium using the usual H&E staining. This will facilitate the work of the pathologist by providing a fast, easy-to-use tool that creates the premises of reproducibility in the quantification of fibrosis. The method is based on the CIELAB color space color variation analysis algorithm implemented in MATLAB® version R2024a and Python version 3.11.7. The implementation in MATLAB® started from results available in [27] and has been improved to work efficiently on histopathologic images, while the implementation in Python belongs to the authors of this paper. The proposed method will be tested by stereology; the preferred method will be used when it is desired to obtain quantitative data from microscopic sections [28], and the quantification will be performed by ImageJ, the most widely used software in this field [29].

2. Materials and Methods

2.1. Database Creation and Image Analysis

Forty digital images of myocardial fibrosis were extracted from the histopathology collection of the digital slide collection of The Iowa Virtual Slidebox [30] using Biolucida Cloud Viewer (version 2024.2.0; MBF Bioscience-MicroBrightField, Inc., Williston, VT, USA). Those considered relevant were selected by the pathologist, excluding the subendocardial and subepicardial areas, or the region close to the heart valves. For the qualitative evaluation of the performance of the algorithm, the images obtained immediately after the application of the algorithm were also acquired, later being visually compared by the pathologist with the original ones. Two pathologists manually outlined the areas of myocardial fibrosis to serve as a model for the selection of the region of interest by the researchers who developed and implemented the algorithm.

2.2. Stereology Analysis

Regarding Stereology Analyzer (version 4.3.4; ADCIS, Saint Contest, France), the software superimposes a grid over the image to be analyzed, whose pattern and spacing can be changed. For the present work, we used a sampling interval of 30 pixels and a pattern size of 7 pixels, resulting in a minimum of 599 test points for the evaluated images. Each test point was manually evaluated by the pathologist and classified as having fibrosis or not. In the end, the program automatically generates stereology volume fraction (%), in this case CVF, which was calculated as the ratio between the number of test points with fibrosis and the total number of control points.

2.3. ImageJ Analysis

CVF determination by ImageJ, version 1.54g (Wayne Rasband and contributors, National Institutes of Health, Bethesda, MD, USA; http://imagej.org (accessed on 1 August 2024)) was performed manually as follows: after opening the image, converting to grayscale and adjusting contrast and brightness, thresholding values were set to select only areas with fibrosis. Thresholding was performed manually for each image, selecting the “Default” option from the ImageJ menu. The threshold adjustment was performed visually, so the region of interest could be properly highlighted.
For segmentation and measurements, the binary conversion of the image was performed, and through the “Analyze Particles” option, the CVF value was obtained.
Stereology and ImageJ evaluation were performed by the same pathologist 3 months apart, blinded to the results obtained by the other methods.

2.4. Algorithm Description

The developed algorithm is based on the implementation in two programming languages, Python and MATLAB®, of the color analysis method based on the CIELAB spectrum [31]. The multi-language implementation, together with the differentiation of the fibrosis mini-region selection mode, ensures robustness in terms of testing the capabilities of the CIELAB color spectrum in semi-automated fibrosis detection from histopathology images.
In both cases, the first step is to select a histopathological image that the user wishes to analyze. The next step appeals to the user’s knowledge of histopathologic analysis and involves the manual selection of a small part of the histopathologic image that contains fibrosis based on the user’s knowledge. In the MATLAB® implementation, the region has an irregular geometric shape and is determined by the cursor movement, while in Python, the region will always have a rectangular shape.
The next step is performed automatically and involves the color analysis of the region selected by the user. Starting from this color, the developed algorithm analyzes the whole histopathologic image at the pixel level and marks as fibrosis any pixel that has the specific color of the region selected by the user, taking into account the Delta E metric in the CIELAB color space. If the user does not have the necessary skills and selects a mini-region that does not contain fibrosis tissue, the algorithm will generate inappropriate results, detecting tissues in the whole image with a chromaticity similar to the one selected by the user.
The final result is represented as a new image in which the fibrosis tissue is marked using a different color from the original histopathological image. In addition, the performance metrics related to the number of fibrotic pixels, the original image size and the percentage of fibrotic tissue in the entire image are returned to the user in numerical form.
The flow diagram of the system is presented in Figure 1.

2.5. Statistical Analysis

Data were analyzed with the Kolmogorov–Smirnov test to assess the distribution of the variables. Normally distributed results are presented as mean ± standard deviation (SD) or as 95% confidence interval (CI), or as median (25–75% interquartile range [IQR]) if non-normally distributed. The Pearson coefficient (r) was performed to evaluate the correlation between groups, being interpreted as follows: > 0.7, very high correlation; 0.5–0.7, high correlation; 0.3–0.5, moderate correlation; 0.1–0.3, weak correlation; and <0.1, no correlation. Differences between groups were measured using the t-test for paired samples. For multiple comparisons, the Analysis of Variance (ANOVA) was used followed by the Bonferroni post hoc test to assess differences between each pair.
The Bland–Altman analysis was performed to test the concordance between methods. p < 0.05 was considered statistically significant. Online Statistics Calculator (DATAtab e.U. Graz, Austria. URL https://datatab.net (accessed on 1 August 2024)) and MedCalc Statistical Software version 22.032 (MedCalc Software Ltd., Ostend, Belgium; https://www.medcalc.org (accessed on 1 August 2024)) were used for the statistical analysis.

3. Results

Of the 40 images manually extracted from The Iowa Virtual Slidebox [30] with myocardial fibrosis, 1 was excluded because the stereological evaluation was difficult (pale colors, imprecise demarcation between cardiomyocytes and collagen fibers), and another 3 were excluded because the evaluation in ImageJ raised difficulties (poor contrast between fibrosis area and normal cardiomyocytes), resulting in a database of 36 images (n = 36). The means and SDs of CVF obtained by the four methods are shown in Table 1. All data were normally distributed.
When comparing CVF determined using ImageJ with that determined by stereology, a mean difference of 0.57% (95% CI: 1.26–2.4) was observed, with no significant difference between the two groups (p = 0.532). The Pearson correlation coefficient showed that there is a very high, statistically significant correlation between the two (r = 0.88, p < 0.001).

3.1. Performance Evaluation of the Algorithm Implemented in CIELAB-Python

CVF determined by running CIELAB-Python shows a high, positive, and statistically significant correlation, both with stereology (r = 0.51, p = 0.001) and ImageJ (r = 0.56, p < 0.001). Statistically significant differences were observed between CVF, determined on the one hand by CIELAB-Python, and on the other by stereology (p < 0.001) and ImageJ (p < 0.001). The mean differences were 6.13% (95% CI: 2.98–9.34) and 6.7% (95% CI: 4.24–9.17). In Figure 2, the images processed by ImageJ (A) and CIELAB-Python (B) are presented.

3.2. Performance Evaluation of the Algorithm Implemented in CIELAB-MATLAB®

CVF determined by running CIELAB-MATLAB® showed a high, positive, and statistically significant correlation with both stereology (r = 0.64, p < 0.001) and ImageJ (r = 0.6, p < 0.001). Statistically significant differences were observed between CVF, determined on the one hand by CIELAB-MATLAB®, and on the other by stereology (p < 0.001) and ImageJ (p < 0.001). The mean differences were 9.04 (95% CI: 6.22–11.86) and 8.47% (95% CI: 6.49–10.49). Figure 2B highlights the generated layout. The pixels colored in black were those quantified by the algorithm.

3.3. Performance Evaluation of the Combined Algorithm

For further evaluations, the mean CVF between CIELAB-MATLAB® and CIELAB-Python (mean MATLAB®-Python) was calculated for each image.
The differences between the percentages determined by ImageJ, stereology and mean MATLAB®-Python were evaluated with the ANOVA test, which revealed no statistically significant difference between groups (p = 0.45). The results of Bonferroni post hoc tests are shown in Table 2.
The Pearson correlation coefficient showed that there is a high, positive, and statistically significant correlation between mean MATLAB®-Python and ImageJ and stereology (r = 0.59, p < 0.001 and r = 0.58, p < 0.001, respectively), with no significant differences between the two groups (p = 0.39). The results of the Bland–Altman analysis are shown in Table 3 and Figure 3A.
The Pearson correlation coefficient showed that there is a high, positive, and statistically significant correlation between stereology and mean MATLAB®-Python (r = 0.58, p < 0.001), with no significant differences between the two groups (p = 0.318). The results of the Bland–Altman analysis are shown in Figure 3B and Table 4.

4. Discussion

Remarkable advances have been made in the field of medical imaging in recent years thanks to the use of digitalized methods and, more recently, artificial intelligence [32,33].
The problem of validating the existing models is that comparisons are made with the visual assessment performed by the pathologist, which is subject to variability [18]. Stereology allows for the three-dimensional quantitative evaluation of two-dimensional microscopic sections, based on a mathematical principle derived from stochastic geometry [34]. Although a time-consuming method requiring manual analysis of each test point, it remains the gold standard for obtaining quantitative data [28]. Although it was observed that the CVF determined by stereology shows a high correlation with the CVF determined by CIELAB-MATLAB® and with that determined by CIELAB-Python, statistically significant differences were observed between the groups, which is why the performance of the combined algorithm was further tested. The CVF, calculated as the average of the values generated by the two algorithms, was not significantly different from the CVF calculated by stereology. Considering that any quantitative assessment involves error, the Bland–Altman analysis aims to establish the degree of agreement between two quantitative methods, allowing comparability [35]. Besides the quantification of mean bias, the method involves the expression of the limits of agreement (LoA), an interval in which 95% of the measured differences are included [35]. The Bland–Altman analysis revealed a relatively small mean of the differences between the two methods (1.5%), which was not significant, considering that the line of equality was included in the 95% CI of the mean difference. However, the LoA were relatively large (95% LoA—15.4–18.3).
Many image analysis programs have been developed over time, but ImageJ remains the most widely used [36], which is the reason why it was chosen as a second method for the validation of the results. A disadvantage of ImageJ is that low-contrast tissues can be difficult to assess [29], therefore three of the images were excluded from the study. A high correlation was also observed between the CVF determined by ImageJ and CIELAB-MATLAB® and CIELAB-Python, respectively, but still with significant differences between groups. The combined algorithm, however, showed a superior performance, with a mean bias of 0.9%, which was not significant (95% CI—1.2–2.9), and a narrower LoA than in the case of the comparison with stereology (95% LoA—11.2–13). Although the two proposed methods (MATLAB®, Python) quantify the same parameter, they offer different results because they are based on different approaches. The fibrosis mini-region is manually selected by the user, which induces randomness based on expertise. Furthermore, in MATLAB®, the selected region can have any shape, while in Python the selected region will always be a rectangle. In addition to the different working principles of the evaluated methods, the significant differences observed between the groups can also be caused by the management of the artifacts. In stereology, each test point is approached individually by the pathologist, while the algorithms implemented in Python or MATLAB® decide if a pixel represents fibrosis or not, depending on the color analysis.
Computer-assisted quantification of myocardial fibrosis is not a new idea, but it is an under-investigated field. A semi-automatic method implemented in ImageJ, also based on the CIELAB system reported by Hadi AM et al. [37], shows a very high correlation with stereology. The method was tested on post-mortem myocardial tissue stained with picrosirius red. The researchers reported a significant reduction in the time needed for evaluation (22% of the time needed for stereology). Stereology was also used as a reference in the study performed by Daunoravicius D et al. [18], which tested the assessment of fibrosis on endomyocardial biopsies. The authors proposed two different algorithms, based on colocalization and pattern recognition, both with high correlation with stereology, but with a minimal underestimation of the second one. The mean bias reported was close to zero, with the LoA was around ±10%. In the study by Vasiljević JD et al. [38], the commercially available software tested, which identifies fibrosis based on color analysis, overestimated CVF by 3 ± 6.7%, with LoA of up to 16.3%. The LoA identified in the present study for the combined method, compared to stereology, is in line with those mentioned in the previous studies. Another method proposed by Zimmermann E et al. [39] for the quantification of myocardial fibrosis assumed infrared spectroscopic imaging combined with machine learning algorithms, and was able to generate virtual images stained with picrosirius red. Multiphoton microscopy has also proven its valuable contribution in this field, being able to visualize the microscopic structure of the myocardium on unstained sections [40].
The aim of this study was to validate a facile, semi-automatic method to quantify myocardial fibrosis and to compare it with stereological and ImageJ evaluation. In the case of this paper, the manual identification of fibrosis using histopathologic images was performed by two pathologists, and the user of the software application selected the mini-region, taking into account the manual markings. Moderate performances were observed for the proposed methods; however, performance was superior in the case of the combined method. Another limitation was caused by the examination of digital images, and not of the slide as a whole, which is why the calculated percentages cannot refer to the entire tissue fragment [41]. In a real-world scenario, there is a possibility of over- or underestimation of fibrosis [42,43]. The semi-automatic character of the algorithm is also subjected to intra- and inter-observer variability, as the ROI must be selected by an expert in the field that can distinguish between normality and abnormality. Thereby, the algorithm can be an aid to the pathologist, not a substitute. In a real-world scenario, the pathologist will only need to identify a mini-region with fibrosis in a histopathological image, and the algorithm will automatically identify fibrosis in the whole image, calculate the percentage of fibrosis tissue in the whole image and represent this fibrosis tissue with a different color. All these features reduce the time needed for pathologists to manually evaluate histopathological images to detect fibrosis in real life.
The differences observed between the methods can also be caused by the fact that a special stain for collagen was not used, a fact that predisposes results to poor contrast [37]. Among the variety of staining methods for collagen emphasis, it appears that the most suitable is picrosirius red, combined with subsequent examination in polarized light [44]. It also has an affinity for the fine collagen type III fibers, unlike Van Gieson or Masson trichrome stainings [44]. Therefore, the results presented may be underestimated. Regardless of the stain chosen, color variation is a problem in digital pathology methods, being attributed to the staining (different staining conditions, different reagents) and scanning slides of under different conditions [45]. Significant qualitative differences have been reported, even among slides stained in the same laboratory [16]. An advantage of the methods analyzed in the present study is that they showed a strong correlation with ImageJ quantification and stereology, although the database included diverse images in terms of color variability, contrast and artifacts encountered. New deep learning methods can generate virtual images with certain special stains, starting from H&E [10], with the potential to generate more comparable images. It is also important to consider the thickness of the sections because the area of collagen deposition is proportional to it [42].
The very high correlation observed between CVF determined by stereology and that assessed by ImageJ shows us that, in the conditions of exposed variability, judicious examination by the pathologist is the one that receives and provides reproducible results, an observation that was also reported by Astbury et al. [16]. In the present paper, the thresholding in ImageJ was performed manually because it was reported to be more accurate compared to automatic or semi-automatic segmentation, even though is more time-consuming [46]. Even so, considering the manual nature of this stage, it is a source of error due to subjectivity [47]. Multiple automatic methods have been reported, but choosing one is still a personal decision [48].
The authors of the present paper are, however, aware of this study’s limitations. The presented data represent a preliminary pilot study on the accuracy of CIELAB-Python and CIELAB-MATLAB®. In order to achieve a power of 80% and a level of significance of 5% (two-sided), for detecting a true difference in means between the test and the standard reference group, the authors calculated a required sample size of 97 images.

5. Conclusions

The combination of the two proposed methods (CIELAB-Python and CIELAB-MATLAB®) has the potential to provide pathologists with a semi-automated method for quantifying myocardial fibrosis on H&E-stained slides. The algorithm was compared with Stereology and ImageJ, being superior due to its ability to reduce the time dedicated to analysis. Although the differences between the methods were small, the limits of agreement were relatively large, reflecting the variability of the images included in the database. Even though the obtained results are encouraging, a retest of the method should be taken into consideration to eliminate variable factors, especially those caused by the stain, on a larger data set, including the algorithm’s performance across multiple users or with different stains. Moreover, regarding future directions, considering the generality of the proposed method in this paper for color analysis and marking pixels having a color close to a selected mini-region, the algorithm may be tested on images with fibrosis with a different localization (liver, kidney) or even on different lesions involving chromaticity differences (calcifications, amyloid deposits), requiring only the change of the region selected by the user.

Author Contributions

Conceptualization, L.A.-C., E.-H.D. and T.M.; methodology, D.G., S.-M.N. and A.-G.B.; software, A.-G.B., A.E.D. and N.L.; validation, D.G. and A.-G.B.; writing—original draft preparation, D.G.; writing—review and editing, A.-G.B., A.E.D., N.L., S.-M.N., T.M., E.-H.D. and L.A.-C.; supervision L.A.-C. and E.-H.D. All authors have read and agreed to the published version of the manuscript.

Funding

This work was partially supported by a grant of the Ministry of Research, Innovation and Digitization, CNCS-UEFISCDI, project number PN-III-P4-PCE-2021-0750, within PNCDI III. Alexandru G. Berciu acknowledges project 38 PFE in the frame of the program PDI-PFE-CDI 2021.

Institutional Review Board Statement

Not applicable.

Informed Consent Statement

Patient consent was waived due to using only images extracted from the open access database (The Iowa Virtual Slidebox).

Data Availability Statement

The original contributions presented in the study are included in the article, further inquiries can be directed to the corresponding authors.

Acknowledgments

We acknowledge the contribution of the Iowa Virtual Slidebox, an open-access database, through which the microscopic images from the present study could be obtained.

Conflicts of Interest

The authors declare no conflicts of interest.

References

  1. Yancy, C.W.; Jessup, M.; Bozkurt, B.; Butler, J.; Casey, D.E.; Drazner, M.H.; Fonarow, G.C.; Geraci, S.A.; Horwich, T.; Januzzi, J.L.; et al. 2013 ACCF/AHA Guideline for the Management of Heart Failure: Executive Summary. Circulation 2013, 128, 1810–1852. [Google Scholar] [CrossRef] [PubMed]
  2. Schirone, L.; Forte, M.; Palmerio, S.; Yee, D.; Nocella, C.; Angelini, F.; Pagano, F.; Schiavon, S.; Bordin, A.; Carrizzo, A.; et al. A Review of the Molecular Mechanisms Underlying the Development and Progression of Cardiac Remodeling. Oxid. Med. Cell Longev. 2017, 2017, 3920195. [Google Scholar] [CrossRef]
  3. Azevedo, P.S.; Polegato, B.F.; Minicucci, M.F.; Paiva, S.A.R.; Zornoff, L.A.M. Cardiac Remodeling: Concepts, Clinical Impact, Pathophysiological Mechanisms and Pharmacologic Treatment. Arq. Bras. Cardiol. 2016, 106, 62–69. [Google Scholar] [CrossRef]
  4. Nauffal, V.; Di Achille, P.; Klarqvist, M.D.R.; Cunningham, J.W.; Hill, M.C.; Pirruccello, J.P.; Weng, L.-C.; Morrill, V.N.; Choi, S.H.; Khurshid, S.; et al. Genetics of Myocardial Interstitial Fibrosis in the Human Heart and Association with Disease. Nat. Genet. 2023, 55, 777–786. [Google Scholar] [CrossRef]
  5. Liu, T.; Song, D.; Dong, J.; Zhu, P.; Liu, J.; Liu, W.; Ma, X.; Zhao, L.; Ling, S. Current Understanding of the Pathophysiology of Myocardial Fibrosis and Its Quantitative Assessment in Heart Failure. Front. Physiol. 2017, 8, 238. [Google Scholar] [CrossRef]
  6. Bengel, F.M.; Diekmann, J.; Hess, A.; Jerosch-Herold, M. Myocardial Fibrosis: Emerging Target for Cardiac Molecular Imaging and Opportunity for Image-Guided Therapy. J. Nucl. Med. 2023, 64, S49–S58. [Google Scholar] [CrossRef]
  7. Galati, G.; Leone, O.; Pasquale, F.; Olivotto, I.; Biagini, E.; Grigioni, F.; Pilato, E.; Lorenzini, M.; Corti, B.; Foà, A.; et al. Histological and Histometric Characterization of Myocardial Fibrosis in End-Stage Hypertrophic Cardiomyopathy. Circ. Heart Fail. 2016, 9, e003090. [Google Scholar] [CrossRef]
  8. López, B.; Ravassa, S.; Moreno, M.U.; José, G.S.; Beaumont, J.; González, A.; Díez, J. Diffuse Myocardial Fibrosis: Mechanisms, Diagnosis and Therapeutic Approaches. Nat. Rev. Cardiol. 2021, 18, 479–498. [Google Scholar] [CrossRef] [PubMed]
  9. Arteaga, E.; de Araújo, A.Q.; Bernstein, M.; Ramires, F.J.A.; Ianni, B.M.; Fernandes, F.; Mady, C. Prognostic Value of the Collagen Volume Fraction in Hypertrophic Cardiomyopathy. Arq. Bras. Cardiol. 2009, 92, 210–214, 216–220. [Google Scholar] [CrossRef]
  10. Naglah, A.; Khalifa, F.; El-Baz, A.; Gondim, D. Conditional GANs Based System for Fibrosis Detection and Quantification in Hematoxylin and Eosin Whole Slide Images. Med. Image Anal. 2022, 81, 102537. [Google Scholar] [CrossRef]
  11. Farris, A.B.; Vizcarra, J.; Amgad, M.; Donald Cooper, L.A.; Gutman, D.; Hogan, J. Image Analysis Pipeline for Renal Allograft Evaluation and Fibrosis Quantification. Kidney Int. Rep. 2021, 6, 1878–1887. [Google Scholar] [CrossRef] [PubMed]
  12. Bera, K.; Schalper, K.A.; Rimm, D.L.; Velcheti, V.; Madabhushi, A. Artificial Intelligence in Digital Pathology—New Tools for Diagnosis and Precision Oncology. Nat. Rev. Clin. Oncol. 2019, 16, 703–715. [Google Scholar] [CrossRef] [PubMed]
  13. Dulf, E.-H.; Bledea, M.; Mocan, T.; Mocan, L. Automatic Detection of Colorectal Polyps Using Transfer Learning. Sensors 2021, 21, 5704. [Google Scholar] [CrossRef]
  14. Stoleru, C.-A.; Dulf, E.H.; Ciobanu, L. Automated Detection of Celiac Disease Using Machine Learning Algorithms. Sci. Rep. 2022, 12, 4071. [Google Scholar] [CrossRef] [PubMed]
  15. Masseroli, M.; Caballero, T.; O’Valle, F.; Del Moral, R.M.; Pérez-Milena, A.; Del Moral, R.G. Automatic Quantification of Liver Fibrosis: Design and Validation of a New Image Analysis Method: Comparison with Semi-Quantitative Indexes of Fibrosis. J. Hepatol. 2000, 32, 453–464. [Google Scholar] [CrossRef]
  16. Astbury, S.; Grove, J.I.; Dorward, D.A.; Guha, I.N.; Fallowfield, J.A.; Kendall, T.J. Reliable Computational Quantification of Liver Fibrosis Is Compromised by Inherent Staining Variation. J. Pathol. Clin. Res. 2021, 7, 471–481. [Google Scholar] [CrossRef] [PubMed]
  17. Sánchez-Jaramillo, E.A.; Gasca-Lozano, L.E.; Vera-Cruz, J.M.; Hernández-Ortega, L.D.; Salazar-Montes, A.M. Automated Computer-Assisted Image Analysis for the Fast Quantification of Kidney Fibrosis. Biology 2022, 11, 1227. [Google Scholar] [CrossRef]
  18. Daunoravicius, D.; Besusparis, J.; Zurauskas, E.; Laurinaviciene, A.; Bironaite, D.; Pankuweit, S.; Plancoulaine, B.; Herlin, P.; Bogomolovas, J.; Grabauskiene, V.; et al. Quantification of Myocardial Fibrosis by Digital Image Analysis and Interactive Stereology. Diagn. Pathol. 2014, 9, 114. [Google Scholar] [CrossRef]
  19. Wang, T.; Rentería, M.E.; Peng, J. Editorial: Data Mining and Statistical Methods for Knowledge Discovery in Diseases Based on Multimodal Omics. Front. Genet. 2022, 13, 895796. [Google Scholar] [CrossRef]
  20. Reel, P.S.; Reel, S.; Pearson, E.; Trucco, E.; Jefferson, E. Using Machine Learning Approaches for Multi-Omics Data Analysis: A Review. Biotechnol. Adv. 2021, 49, 107739. [Google Scholar] [CrossRef]
  21. Li, Y.-H.; Cao, Y.; Liu, F.; Zhao, Q.; Adi, D.; Huo, Q.; Liu, Z.; Luo, J.-Y.; Fang, B.-B.; Tian, T.; et al. Visualization and Analysis of Gene Expression in Stanford Type A Aortic Dissection Tissue Section by Spatial Transcriptomics. Front. Genet. 2021, 12, 698124. [Google Scholar] [CrossRef]
  22. Wang, T.; Shu, H.; Hu, J.; Wang, Y.; Chen, J.; Peng, J.; Shang, X. Accurately Deciphering Spatial Domains for Spatially Resolved Transcriptomics with stCluster. Brief. Bioinform. 2024, 25, bbae329. [Google Scholar] [CrossRef]
  23. Li, W.; Lou, X.; Zha, Y.; Qin, Y.; Zha, J.; Hong, L.; Xie, Z.; Yang, S.; Wang, C.; An, J.; et al. Single-Cell RNA-Seq of Heart Reveals Intercellular Communication Drivers of Myocardial Fibrosis in Diabetic Cardiomyopathy. Elife 2023, 12, e80479. [Google Scholar] [CrossRef] [PubMed]
  24. Long, X.; Yuan, X.; Du, J. Single-Cell and Spatial Transcriptomics: Advances in Heart Development and Disease Applications. Comput. Struct. Biotechnol. J. 2023, 21, 2717–2731. [Google Scholar] [CrossRef]
  25. Yang, M.Q.; Weissman, S.M.; Yang, W.; Zhang, J.; Canaann, A.; Guan, R. MISC: Missing Imputation for Single-Cell RNA Sequencing Data. BMC Syst. Biol. 2018, 12, 114. [Google Scholar] [CrossRef] [PubMed]
  26. Wang, T.; Zhao, H.; Xu, Y.; Wang, Y.; Shang, X.; Peng, J.; Xiao, B. scMultiGAN: Cell-Specific Imputation for Single-Cell Transcriptomes with Multiple Deep Generative Adversarial Networks. Brief. Bioinform. 2023, 24, bbad384. [Google Scholar] [CrossRef]
  27. Color Segmentation by Delta E Color Difference. Available online: https://www.mathworks.com/matlabcentral/fileexchange/31118-color-segmentation-by-delta-e-color-difference (accessed on 22 August 2024).
  28. Mühlfeld, C.; Schipke, J. Methodological Progress of Stereology in Cardiac Research and Its Application to Normal and Pathological Heart Development. Cells 2022, 11, 2032. [Google Scholar] [CrossRef]
  29. Gratz, D.; Winkle, A.J.; Dalic, A.; Unudurthi, S.D.; Hund, T.J. Computational Tools for Automated Histological Image Analysis and Quantification in Cardiac Tissue. Methods X 2020, 7, 100755. [Google Scholar] [CrossRef]
  30. Dee, F.R.; Leaven, T. Iowa Virtual Slidebox; University of Iowa Healthcare: Iowa City, IA, USA, 1999; Available online: https://biolucida.net/viewer/?page=viewer (accessed on 27 August 2024).
  31. Cristoforetti, A.; Masè, M.; Ravelli, F. Model-Based Approach for the Semi-Automatic Analysis of Collagen Birefringence in Polarized Light Microscopy. Appl. Sci. 2023, 13, 2916. [Google Scholar] [CrossRef]
  32. Schüffler, P.; Steiger, K.; Weichert, W. How to Use AI in Pathology. Genes. Chromosomes Cancer 2023, 62, 564–567. [Google Scholar] [CrossRef]
  33. Danku, A.E.; Dulf, E.H.; Banut, R.P.; Silaghi, H.; Silaghi, C.A. Cancer Diagnosis with the Aid of Artificial Intelligence Modeling Tools. IEEE Access 2022, 10, 20816–20831. [Google Scholar] [CrossRef]
  34. Knudsen, L.; Brandenberger, C.; Ochs, M. Stereology as the 3D Tool to Quantitate Lung Architecture. Histochem. Cell Biol. 2021, 155, 163–181. [Google Scholar] [CrossRef]
  35. Giavarina, D. Understanding Bland Altman Analysis. Biochem. Med. 2015, 25, 141–151. [Google Scholar] [CrossRef]
  36. Schneider, C.A.; Rasband, W.S.; Eliceiri, K.W. NIH Image to ImageJ: 25 Years of Image Analysis. Nat. Methods 2012, 9, 671–675. [Google Scholar] [CrossRef] [PubMed]
  37. Hadi, A.M.; Mouchaers, K.T.B.; Schalij, I.; Grunberg, K.; Meijer, G.A.; Vonk-Noordegraaf, A.; van der Laarse, W.J.; Beliën, J.A.M. Rapid Quantification of Myocardial Fibrosis: A New Macro-Based Automated Analysis. Anal. Cell. Pathol. 2010, 33, 257–269. [Google Scholar] [CrossRef]
  38. Vasiljević, J.D.; Popović, Z.B.; Otasević, P.; Popović, Z.V.; Vidaković, R.; Mirić, M.; Nesković, A.N. Myocardial Fibrosis Assessment by Semiquantitative, Point-Counting and Computer-Based Methods in Patients with Heart Muscle Disease: A Comparative Study. Histopathology 2001, 38, 338–343. [Google Scholar] [CrossRef] [PubMed]
  39. Zimmermann, E.; Mukherjee, S.S.; Falahkheirkhah, K.; Gryka, M.C.; Kajdacsy-Balla, A.; Hasan, W.; Giraud, G.; Tibayan, F.; Raman, J.; Bhargava, R. Detection and Quantification of Myocardial Fibrosis Using Stain-Free Infrared Spectroscopic Imaging. Arch. Pathol. Lab. Med. 2021, 145, 1526–1535. [Google Scholar] [CrossRef] [PubMed]
  40. Yang, Y.; Zheng, L.; Li, Z.; Chen, J.; Wu, X.; Ren, G.; Xiao, Z.; Li, X.; Luo, W.; Wu, Z.; et al. Multiphoton Microscopy Providing Pathological-Level Quantification of Myocardial Fibrosis in Transplanted Human Heart. Lasers Med. Sci. 2022, 37, 2889–2898. [Google Scholar] [CrossRef]
  41. Hsia, C.C.W.; Hyde, D.M.; Ochs, M.; Weibel, E.R. ATS/ERS Joint Task Force on Quantitative Assessment of Lung Structure An Official Research Policy Statement of the American Thoracic Society/European Respiratory Society: Standards for Quantitative Assessment of Lung Structure. Am. J. Respir. Crit. Care Med. 2010, 181, 394–418. [Google Scholar] [CrossRef]
  42. Schipke, J.; Brandenberger, C.; Rajces, A.; Manninger, M.; Alogna, A.; Post, H.; Mühlfeld, C. Assessment of Cardiac Fibrosis: A Morphometric Method Comparison for Collagen Quantification. J. Appl. Physiol. 2017, 122, 1019–1030. [Google Scholar] [CrossRef]
  43. Testa, L.C.; Jule, Y.; Lundh, L.; Bertotti, K.; Merideth, M.A.; O’Brien, K.J.; Nathan, S.D.; Venuto, D.C.; El-Chemaly, S.; Malicdan, M.C.V.; et al. Automated Digital Quantification of Pulmonary Fibrosis in Human Histopathology Specimens. Front. Med. 2021, 8, 607720. [Google Scholar] [CrossRef] [PubMed]
  44. López De Padilla, C.M.; Coenen, M.J.; Tovar, A.; De la Vega, R.E.; Evans, C.H.; Müller, S.A. Picrosirius Red Staining: Revisiting Its Application to the Qualitative and Quantitative Assessment of Collagen Type I and Type III in Tendon. J. Histochem. Cytochem. 2021, 69, 633–643. [Google Scholar] [CrossRef] [PubMed]
  45. Murakami, Y.; Abe, T.; Hashiguchi, A.; Yamaguchi, M.; Saito, A.; Sakamoto, M. Color Correction for Automatic Fibrosis Quantification in Liver Biopsy Specimens. J. Pathol. Inform. 2013, 4, 36. [Google Scholar] [CrossRef] [PubMed]
  46. Hore, S.; Chakraborty, S.; Chatterjee, S.; Dey, N.; Ashour, A.S.; Chung, L.V.; Le, D.-N. An Integrated Interactive Technique for Image Segmentation Using Stack Based Seeded Region Growing and Thresholding. IJECE 2016, 6, 2773. [Google Scholar] [CrossRef]
  47. Moreira, A.C.; Appoloni, C.R.; Mantovani, I.F.; Fernandes, J.S.; Marques, L.C.; Nagata, R.; Fernandes, C.P. Effects of Manual Threshold Setting on Image Analysis Results of a Sandstone Sample Structural Characterization by X-Ray Microtomography. Appl. Radiat. Isot. 2012, 70, 937–941. [Google Scholar] [CrossRef]
  48. Tadrous, P.J. On the Concept of Objectivity in Digital Image Analysis in Pathology. Pathology 2010, 42, 207–211. [Google Scholar] [CrossRef]
Applsci 14 07696 g001
Figure 1. The flow diagram of the system.
Figure 1. The flow diagram of the system.
Applsci 14 07696 g001
Applsci 14 07696 g002
Figure 2. (A) Binary image after processing in ImageJ (fibrosis is represented in white). (B) CIELAB-MATLAB® (fibrosis is represented in pink). (C) CIELAB-Python (fibrosis is represented in black). Original image: The Iowa Virtual Slidebox [30].
Figure 2. (A) Binary image after processing in ImageJ (fibrosis is represented in white). (B) CIELAB-MATLAB® (fibrosis is represented in pink). (C) CIELAB-Python (fibrosis is represented in black). Original image: The Iowa Virtual Slidebox [30].
Applsci 14 07696 g002
Applsci 14 07696 g003
Figure 3. Bland–Altman plot based on the percentages of fibrosis returned by the mean MATLAB®-Python and ImageJ (A) or stereology (B). The differences between the two paired methods are plotted against the mean of the measurements. Mean bias (solid blue line) with 95% CI (illustrated in green) and 95% limits of agreement (dashed red lines) with their CI are marked.
Figure 3. Bland–Altman plot based on the percentages of fibrosis returned by the mean MATLAB®-Python and ImageJ (A) or stereology (B). The differences between the two paired methods are plotted against the mean of the measurements. Mean bias (solid blue line) with 95% CI (illustrated in green) and 95% limits of agreement (dashed red lines) with their CI are marked.
Applsci 14 07696 g003
Table 1. Descriptive parameters of the percentages of fibrosis assessed by different methods.
Table 1. Descriptive parameters of the percentages of fibrosis assessed by different methods.
ImageJ (%)Stereology (%)CIELAB-Python (%)CIELAB-MATLAB® (%)
n36363636
Mean ± SD29.76 ± 7.2530.33 ± 10.5636.46 ± 8.1121.29 ± 5.01
SD standard deviation.
Table 2. Bonferroni post hoc tests.
Table 2. Bonferroni post hoc tests.
Mean Diff.Std. Errortp95% CI Lower Limit95% CI Upper Limit
Stereology (%)ImageJ (%)0.570.9010.6311−1.262.4
Stereology (%)Mean MATLAB®-Python1.451.4341.0130.954−1.464.37
ImageJ (%)Mean MATLAB®-Python0.881.0310.8581−1.212.98
CI confidence interval.
Table 3. Summary of Bland–Altman analysis of the percentages of fibrosis returned by the mean MATLAB®-Python and ImageJ.
Table 3. Summary of Bland–Altman analysis of the percentages of fibrosis returned by the mean MATLAB®-Python and ImageJ.
Summary of Bland–Altman
Arithmetic mean0.8847
95% Confidence interval−1.2087 to 2.9782
P (H0: Mean = 0)0.3968
Lower limit−11.2423
95% Confidence interval−14.8535 to −7.6310
Upper limit13.0117
95% Confidence interval9.4005 to 16.6230
Table 4. Summary of Bland–Altman analysis of percentages of fibrosis returned by mean MATLAB®-Python and stereology.
Table 4. Summary of Bland–Altman analysis of percentages of fibrosis returned by mean MATLAB®-Python and stereology.
Summary of Bland–Altman
Arithmetic mean1.4531
95% Confidence interval−1.4582 to 4.3643
P (H0: Mean = 0)0.3179
Lower limit−15.4111
95% Confidence interval−20.4330 to −10.3892
Upper limit18.3172
95% Confidence interval23.3391
Disclaimer/Publisher’s Note: The statements, opinions and data contained in all publications are solely those of the individual author(s) and contributor(s) and not of MDPI and/or the editor(s). MDPI and/or the editor(s) disclaim responsibility for any injury to people or property resulting from any ideas, methods, instructions or products referred to in the content.

Share and Cite

MDPI and ACS Style

Gonciar, D.; Berciu, A.-G.; Danku, A.E.; Lorenzovici, N.; Dulf, E.-H.; Mocan, T.; Nicula, S.-M.; Agoston-Coldea, L. The Quantification of Myocardial Fibrosis on Human Histopathology Images by a Semi-Automatic Algorithm. Appl. Sci. 2024, 14, 7696. https://doi.org/10.3390/app14177696

AMA Style

Gonciar D, Berciu A-G, Danku AE, Lorenzovici N, Dulf E-H, Mocan T, Nicula S-M, Agoston-Coldea L. The Quantification of Myocardial Fibrosis on Human Histopathology Images by a Semi-Automatic Algorithm. Applied Sciences. 2024; 14(17):7696. https://doi.org/10.3390/app14177696

Chicago/Turabian Style

Gonciar, Diana, Alexandru-George Berciu, Alex Ede Danku, Noemi Lorenzovici, Eva-Henrietta Dulf, Teodora Mocan, Sorina-Melinda Nicula, and Lucia Agoston-Coldea. 2024. "The Quantification of Myocardial Fibrosis on Human Histopathology Images by a Semi-Automatic Algorithm" Applied Sciences 14, no. 17: 7696. https://doi.org/10.3390/app14177696

Note that from the first issue of 2016, this journal uses article numbers instead of page numbers. See further details here.

Article Metrics

Back to TopTop