1. Introduction
Hepatocellular carcinomas (HCC) constitute up to 80% of all primary liver cancers [
1]. Liver cirrhosis—often induced by viral hepatitis B and C, alcohol consumption, chemical toxins, or metabolic liver diseases—precedes HCC in at least 80% of cases [
2]. Furthermore, the global rise of obesity and type 2 diabetes leads to the increased incidence of non-alcoholic fatty liver disease (NAFLD), a primary cause of HCC in the absence of cirrhosis [
3]. In recent decades, the impact of HCC has grown significantly, now ranking as the third leading cause of cancer-related mortality worldwide [
4]. From 1990 to 2019, there was a staggering 70% increase in HCC incidence, resulting in 480,000 attributable deaths, exacerbated by late diagnoses and advanced disease stages [
1,
5]. While early detection leads to a five-year survival rate of over 70%, the rate plummets to a mere 18% in later stages [
4,
6]. These trends highlight the necessity for research into surgical techniques, pharmaceutical options, and prognostic factors to meet the challenge posed by HCC.
Unlike many other cancers, post-resection clinical outcomes of HCC patients do not rely solely on tumor properties and success of its therapy. An important additional determinant is the pathology and functional capacity of the residual liver tissue, underscoring the need for comprehensive assessment of both tumor and non-tumor components [
7]. This aspect was taken into account by the Barcelona Clinic Liver Cancer (BCLC) staging and treatment strategy. BLCL points to the limitations of scoring systems for liver failure, like the Child–Pugh system and the model for end-stage liver disease (MELD), which fail to accurately predict the loss of liver function specifically in the context of HCC [
8]. To address this, BCLC staging incorporates multiple parameters of tumor burden, liver function, and cancer-related symptoms, as well as alpha-fetoprotein (AFP) levels and the albumin–bilirubin score for liver function assessment [
9].
Surgical liver resection provides samples containing HCC and liver parenchyma, as well as their interface. This opens multiple opportunities of tissue pathology methods to assess cellular, molecular, and architectural properties of the disease in the spatial context of the tissue microenvironment. A particular aspect that unifies both tumor and non-tumor tissues is represented by properties of the extracellular matrix (ECM), which has been shown to provide rich and quantifiable data of clinical significance. Tumor associated collagen signatures (TACS) have been conceptualized as computational biomarkers to reflect changes in the organization, alignment, and composition of fibers in the stroma that surround and interact with cancer cells [
10]. Originally proposed and demonstrated to have prognostic significance in breast cancer tissue by Keely et al., collagen-derived features were subsequently refined and described in oral, gastric, salivary gland, skin neoplasms and benign fibrous lesions [
11,
12,
13,
14]. On the other hand, microarchitectural transformations of liver tissue represent progression of chronic liver disease.
At the molecular level, the ECM is represented by co-polymers comprising various types of collagen, non-collagenous glycoproteins, proteoglycans, and other molecules. These composite biological materials have distinct characteristics, and thus bear similarities with metal alloys, as noted by Bruckner [
15]. Arranged into fibrillary structures of the interstitial matrix and the basement membrane—the main components of the ECM—they provide a structural foundation for the liver parenchyma [
10]. The molecular complexity of ECM became evident in the 1970s with the discovery that normal liver tissue primarily contains three types of collagen: Type I, Type III, and basement membrane collagens. Type III collagen is the main constituent of reticulin fibers [
16]. However, persistent liver injury leads to significant alterations of the composition, orientation, and quantity of all collagen types [
17]. As fibrosis progresses, Type I collagen accumulates, eventually becoming predominant in the cirrhotic liver [
18]. Despite well-established association between cirrhosis and hepatocarcinogenesis, the exact role of Type I collagen remains unclear: some studies associate it with HCC progression [
19], whereas others propose that Type I collagen accumulation may have a beneficial effect by mechanically restraining tumor spread [
20]. Conversely, the role of Type III collagen is more established, given that the progressive distortion and dissolution of reticulin framework are histopathological hallmarks of HCC [
21]. Therefore, a comprehensive analysis of Type I and Type III (reticulin) collagen properties could provide insights for clinical assessment of patients with HCC.
Currently, histochemical or immunohistochemical methods are used to highlight the different types of collagen [
22]. Assessment of the fiber microarchitecture can be performed by visual inspection of the patterns; however, rapid development of computational methods in digital pathology brings novel opportunities for high-capacity quantification of the structural patterns [
23]. It has been shown that AI solutions are capable of extracting subvisual features with prognostic relevance from liver tissue [
24]. Recently, Patil et al. developed a deep learning model for quantifying reticulin (represented by black, silver-impregnated fibers) in HCC tissue after liver resection [
25]. They found that a decreased reticulin proportionate area (RPA) was an independent predictor of metastasis, shorter disease-free survival, and worse overall survival in this study. Similarly, Taylor-Weiner et al. demonstrated the utility of a convolutional neural network (CNN) for Ishak and NASH Clinical Research Network fibrosis scoring in trichrome-stained slides [
26]. A recent study suggests that liver pathologists are eager for further development of digital pathology and AI integration [
27].
Morkunas et al. proposed a more detailed method for extracting collagen-derived prognostic features; they used a CNN to segment collagen from images of tissue microarrays (TMA) containing Picrosirius Red stained samples of ductal breast carcinoma [
28]. From 37 features of fiber morphometry, density, orientation, texture, and fractal characteristics, they found an independent prognostic value of observed heterogeneity of distances between collagen fibers, fiber straightness, and variance of fiber orientation angles to predict patient survival, even though their method was limited to samples of small amount of tumor tissue (TMA cores). On the other hand, full-face surgical resection samples contain large amount of data that are also affected by intratissue heterogeneity, including the malignant and non-malignant components and their interfaces. Therefore, it is essential to properly assess the spatial aspects of pathology features extracted. To tackle this complexity, Plancoulaine et al. proposed a hexagonal tiling approach that allowed quantification of intratumoral heterogeneity of biomarker expression in breast cancer [
29]. Building upon this method, a tool for automatic detection of the tumor-stroma interface was further developed by Rasmusson et al. [
30].
In this study, we explored the predictive value of linear reticulin and thick septal collagen fibers in both HCC and adjacent liver. Utilizing AI for fiber segmentation and tissue classification, followed by hexagonal grid subsampling of the data, we extracted fiber-specific features in the spatial context of the tissue components and their interfaces. Our findings indicate the potential utility of these features to predict overall survival of patients undergoing liver resection for HCC.
4. Discussion
This study demonstrates the prognostic value of convolutional neural network-based mapping of reticulin and collagen fiber architecture in the HCC microenvironment. The integration of computational features describing the reticulin and collagen texture with the clinical parameters resulted in two multivariate overall survival models with a test cohort C-index > 0.7 after penalized LASSO Cox regression. Both models reveal the independent prognostic impact of patient age, tumor multifocality and fiber-derived features at the interfaces of HCC and the remaining functional hepatocytes with the surrounding stroma. Also, the reticulin structure provided the prognostic value at the tumor edge, while at the border of liver parenchyma, the collagen structure was relevant. Meanwhile, none of the models consisting of conventional clinicopathologic metrics only were able to surpass the 0.7 C-index threshold.
We have discovered that among the HCC-derived features in our cohort, the mean lacunarity of the reticulin framework at the tumor margin was the best-performing metric. Following closely, as indicated by its recurrent appearances in the models (
Table 6), is the variance (SD) of the reticulin lacunarity at the core of the tumor. This observation aligns with the established significance of reticulin which is a key element in the structure of a normal liver, providing the framework of its lobular architecture. The alteration, disruption or dissipation of reticulin fibers is a well-documented diagnostic sign of HCC, first reported nearly 50 years ago [
32]. Lacunarity is a fractal parameter that captures both gaps (Lat. lacuna) and heterogeneity in a pattern. In the context of our study, higher lacunarity would suggest the reduction and distortion of the reticulin framework, possibly indicating a more aggressive tumor phenotype. The high variance of lacunarity at the central part of the tumor—the core—might suggest the existence of areas with different degrees of aggressiveness. This insight finds parallels with a recent study by Patil A. et al., which confirmed a reduction in the AI-identified reticulin proportionate area (RPA) in HCC as a strong predictor of adverse patient outcomes [
25]. However, our work revealed a potential superiority of spatial variations in the reticulin framework, as captured by lacunarity. The underperformance of the mean fiber density (FD, derived from the number of pixels in the mask and comparable to RPA) in our study, in contrast to lacunarity, suggests that the spatial heterogeneity in reticulin arrangement may provide more insight than merely the proportion of reticulin in the HCC tissue.
In cancer diagnostics, the non-neoplastic component of the tissue often receives somewhat lesser attention. However, our findings also underscore the independent prognostic significance of the peritumoral liver parenchyma. The extent of fibrosis, a well-documented predictor of chronic liver disease outcomes, can be assessed using a variety of invasive or non-invasive methods [
33,
34]. We have demonstrated that collagen in the peritumoral liver serves as a significant source of prognostic information, in contrast to the role of reticulin in HCC, as previously discussed. This aligns with the known role of Type I collagen as the primary component of fibrous tissue that accumulates during persistent liver damage. In our study, two features associated with peritumoral liver collagen were included in the most predictive models of overall survival (
Table 5): the mean texture correlation of collagen fibers, and the high variance of collagen fiber straightness. Additionally, the high variance in fiber density emerged as the third collagen-derived feature, listed among the ten most recurrent components in the prognostic models (
Table 6). Importantly, all these indicators were measured at the interface between remaining functional hepatocytes and the fibrous tissue (hexagon ranks −1, 0, 1). The standard deviation of individual measurements serves as an indicator of variability or heterogeneity between hexagons. In this case, it can highlight regions of dense, compact fibrosis in contrast to areas of the liver that remain somewhat intact, characterized by sparsely deposited and less organized collagen fibers. On the other hand, a more consistent overall tissue structure (reflected by the high mean correlation), might indicate a less advanced liver disease with higher residual functional capacity. Combined, the high variability of collagen deposition in the individual hexagons and the maintenance of general parenchymal integrity might indicate the presence of ongoing successful tissue repair. As sufficient residual liver function is crucial for the survival of patients [
35], its assessment alongside tumor parameters on the same tissue slide offers a streamlined and practical approach in HCC resection samples.
A factor analysis was used to investigate the inherent relations between the 26 fiber-derived features that, when combined in certain ways (see
Supplementary Table S2), showed statistically significant predictive power (
p < 0.05) in regression models. Six factors were identified, capturing the majority of the information in the original data. The heatmap in
Figure 5 highlights the strong negative association between the mean reticulin lacunarity at the tumor edge (g_mn_lac_HCC_IZ3) and the dominant Factor 1, positioning it as a key determinant. Notably, this variable also exhibits the most negative loadings for both Factor 2 and Factor 5, and ranks second to last in terms of negative loadings for Factor 4. These consistent negative loadings across four of the six factors suggest that g_mn_lac_HCC_IZ3 captures multidimensional information about the reticulin framework at the tumor edge, making it the most prominent HCC-derived feature. Furthermore, the predominant features defining peritumoral liver-derived Factor 3 reflect the consistency in collagen fiber orientation (r_mn_cor_LVR_IZ3), variability in density (r_sd_FD_LVR_IZ3), and variability in fiber straightness (r_sd_mFS_LVR_IZ3). Consequently, the pairing of a tumor-based g_mn_lac_HCC_IZ3 indicator with either r_mn_cor_LVR_IZ3 or r_sd_mFS_LVR_IZ3, both liver-based features, collectively covers Factors 1–5, which represent 81.04% of the variance. This unique combination demonstrates remarkable performance in Cox regression models and outperforms the factor values, emphasizing the synergistic role of tumor and liver characteristics in HCC prognostication.
Our study contains some limitations. The lack of complete data on HBV and HCV infections restricts a thorough examination of their potential impact on the overall survival in our HCC patients. Secondly, the lack of the information on the cause of death limits our options for predicting disease-specific survival, which would be relevant for our focus on impact of malignant and non-malignant components. Thirdly, the cohort of 105 patients is rather limited and serves as a proof-of-concept study. Validation studies on independent cohorts would be needed to assess generalizability of our findings.