AI-Enhanced Segmentation and Scoring of Histopathological Growth Patterns in Resected Colorectal Liver Metastases
Listed in
This article is not in any list yet, why not save it to one of your lists.Abstract
Histopathological growth patterns, predominantly the presence of an encapsulated rim, serve as powerful prognostic biomarkers in colorectal liver metastases. However, conventional visual estimation is limited by interobserver variability, human rounding bias, and categorical thresholds. We aimed to develop and externally validate a deep-learning segmentation framework that produces continuous, spatially resolved, and objective encapsulation metrics from routine whole-slide images.
In this multicenter study of chemo-naive resected CRLM patients, a modified nnU-Net architecture was developed on 40 pixel-annotated H&E-stained WSIs to segment liver parenchyma, tumor, and the encapsulation interface. Model performance was optimized across spatial resolutions. An automated continuous AI encapsulation score was computed to determine patient-level metrics. The prognostic value for 5-year overall survival was evaluated using restricted cubic spline Cox regression, a log-rank optimized threshold, and Kaplan–Meier survival analyses across a main cohort (Erasmus MC; n = 546, 1,910 WSIs) and an independent external validation cohort (Radboud UMC; n = 212, 615 WSIs).
The segmentation network achieved optimal tissue boundary delineation at a resolution of 2.0 µm/pixel on the 40 annotated slides (overall foreground DICE: 0.86 ± 0.09; encapsulation DICE: 0.70 ± 0.11). Spline modeling demonstrated that while visual assessment displayed non-linear hazard plateaus, the continuous AI encapsulation score exhibited a linear reduction in mortality risk across the critical intermediate range (25%–75% encapsulation, p = 7.81 × 10 −4 ). Log-rank optimization identified a primary global threshold at 47.4% ( p = 3.73 × 10 −8 ), which in confirmed significant OS stratification in external validation ( p = 0.018). A secondary threshold was found at 69% ( p = 2.93 × 10 −6 ) that captured high-encapsulation favorable-risk patients matching the current manual 100% Rotterdam cut-off.
Our deep-learning model transforms qualitative HGP grading into a continuous, explainable, and fully automated prognostic biometry. By eliminating observer subjectivity and capturing fine-grained risk gradients, the model establishes a scalable, standardized foundation for CRLM risk stratification on digital pathology.
