Role of artificial intelligence in staging and assessing of treatment response in MASH patients

Abstract

Background and aims: The risk of disease progression in MASH increases proportionally to the pathological stage of fibrosis. This latter is evaluated through a semi-quantitative process, which has limited sensitivity in reflecting changes in disease or response to treatment. This study aims to test the clinical impact of Artificial Intelligence (AI) in characterizing liver fibrosis in MASH patients.

Methods: The study included 60 patients with clinical pathological diagnosis of MASH. Among these, 17 received a medical treatment and underwent a post-treatment biopsy. For each biopsy (n = 77) a Sirius Red digital slide (SR-WSI) was obtained. AI extracts >30 features from SR-WSI, including estimated collagen area (ECA) and entropy of collagen (EnC).

Results: AI highlighted that different histopathological stages are associated with progressive and significant increase of ECA (F2: 2.6% ± 0.4; F3: 5.7% ± 0.4; F4: 10.9% ± 0.8; p: 0.0001) and EnC (F2: 0.96 ± 0.05; F3: 1.24 ± 0.06; F4: 1.80 ± 0.11, p: 0.0001); disclosed the heterogeneity of fibrosis among pathological homogenous cases; revealed post treatment fibrosis modification in 76% of the cases (vs 56% detected by histopathology).

Conclusion: AI characterizes the fibrosis process by its true, continuous, and non-categorical nature, thus allowing for better identification of the response to anti-MASH treatment.

Keywords: MASH; artificial intelligence; fibrosis; liver; treatment.

Figure 1

Agreement between pathologists in the evaluation of NASH stage. In panel (A), the subgroups of F1, namely F1a, F1b, F1c, have been considered separately, while in panel (B) they have been grouped. The overall Fleiss’s Kappa agreement is 0.44 for panel A and 0.55 for panel B. The agreement for each category is reported over the column. The hypothesis that the agreement is caused by random chance can be rejected (*** p ≤ 0.001).

Figure 2

Correlation between histopathological NASH stage and features of fibrosis generated by AI. The figure illustrates the process of transition from the original image (SR-WSI) to AI-features and the comparison of the histopathological NASH stage (F1 to F4) to a heatmap generated by the AI for ECA and EnC. After preprocessing (see Supplementary Figure S1), the original SR-WSI is analyzed to segment collagen. The quantification process involves the extraction of both intensity and textural features at pixel level and within Regions of Interest (ROIs). Estimated Collagen Area (ECA) is computed as a fraction of collagen pixels (Sirius Red positive) over the total number of pixels representing the tissue section. Entropy of Collagen (EnC) is a textural parameter that encodes for the randomness of SR optical density values with respect to its neighborhood in terms of intensity distribution. Low entropy values correspond to a uniform and homogeneous image.

Figure 3

AI features (ECA and EnC) of single cases according to histopathological NASH stage. The figure shows for each biopsy, grouped according to the histopathological stage, the results of AI evaluation. Panel (A) illustrates the Estimated Collagen Area (ECA), panel (B) describes the evaluation of Entropy of Collagen (EnC). Significant differences between subgroups are assessed by one-way ANOVA corrected by a Tukey post hoc test. (** p ≤ 0.01; *** p ≤ 0.001; **** p ≤ 0.0001). Error bars represent the standard error of mean.

Figure 4

ECA heterogeneity in cases with homogenous histopathological NASH stage. (A,B) SR-WSI of two cases labeled as F2 at histopathological level by all pathologists (perfect agreement). (C,D) Overlay between the original WSI and the Estimated Collagen Area (ECA) heatmap for each case. AI shows that in the first case (heatmap C) the percentage of tissue area covered by collagen (ECA) is 1.67% while in the second case (heatmap D) this percentage is to 3.85%. (E,F) SR-WSI of two cases labeled as F3 at histopathological level, after the adjudication process. (G,H) AI computes, that the area covered by collagen is lower in the first case (heatmap G, ECA = 3.83%) as compared to the second (heatmap H, ECA = 9.95%).

Figure 5

Impact of AI, as compared to histopathology, in the assessment of treatment efficacy taking as reference the post treatment biopsy. The figure compares the results of histopathology and AI in assessing the response to treatment for MASH. Results shown refers to post treatment biopsy (PTB). Histopathology: green = reduction of stage in PTB (responder at histopathology); yellow = no change of stage in PTB (not conclusive at histology); red = increase of stage in PTB (not responder at histopathology). AI (ECA and EnC): green = reduction of value in PTB; red = increase of value in PTB; cases with congrous decrease of ECA and EnC are responders, with congrous decrease not responders, and discordant case as not conclusive, according to AI analysis.

Figure 6

Evaluation of treatment efficacy on fibrosis modification. The figure illustrates the microscopic (SR-WSI, left) and AI (segmentation, ECA and EnC, right) features in pre- and post- treatment biopsy in three patients (A, B, C). In patient A (case 6, Table 2), the pre treatment biopsy (A1) was diagnosed as stage F3, AI reveled a ECA of 9.8% and EnC of 2.04; post treatment biopsy (A2) was still diagnosed as F3 at histopathological level; however AI revealed that EC decreased to 4.8% and EnC to 1.42. In patient B (case 4, Table 2), the pre treatment biopsy (B1) was diagnosed as F3, with ECA of 9.6% and EnC of 1.48; after treatment (B2) the histopathological stage did not change (F3), but AI disclosed a reduction for both ECA (8.0%) and EnC (1.42). In Patient C (case 1, Table 2), the histopathological evaluation disclosed a stage reduction from F4 seen in pre-treatment (C1) to F3 seen in post treatment biopsy (C2); AI features were consistent with this reduction (ECA from 8.6 to 6.3%; EnC from 1.56 to 1.38).