# Multiview deep-learning-enabled histopathology for prognostic and therapeutic stratification in stage II colorectal cancer: A retrospective multicenter study

## TL;DR

A deep learning model called SurvFinder identifies tissue features in colorectal cancer slides to predict relapse risk and guide treatment decisions.

## Contribution

The study introduces SurvFinder, an interpretable deep learning framework that autonomously identifies tissue-based biomarkers for risk stratification in stage II CRC.

## Key findings

- SurvFinder identified tertiary lymphoid structures (TLSs) as critical prognostic features in stage II CRC.
- The model demonstrated superior predictive accuracy across four multicenter datasets with AUROC values ranging from 0.712 to 0.827.
- TLS features such as location and maturity state were found to influence prognosis and adjuvant therapy efficacy.

## Abstract

Approximately 20% of patients with stage II colorectal cancer (CRC) experience tumor relapse despite standard surgical treatment. Histopathological analysis holds promise for postsurgical risk stratification and guiding adjuvant chemotherapy (ACT) decisions. The aim of this study was to use deep learning to extract explainable tissue biomarkers from whole-slide images.

In this retrospective cohort study, we developed and validated SurvFinder, an interpretable deep learning framework designed to autonomously identify tissue-based risk biomarkers from hematoxylin and eosin (H&E)-stained slides. The framework aims to support individualized risk stratification and explore associations with treatment outcomes. The present study included 6,950 H&E slides from 1,604 patients with stage II CRC across four independent cohorts in China. Patients were enrolled from 2012 to 2018 and followed for a minimum of 24 months. The primary outcome of the study was relapse-free survival (RFS). Our analyses identified tertiary lymphoid structures (TLSs) as critical prognostic features in stage II CRC. The multi-view integration of TLS characteristics by SurvFinder consistently demonstrated superior predictive and prognostic accuracy across four multicenter datasets (AUROC with 95% confidence interval [CI]: 0.827 [0.789,0.864], 0.805 [0.749,0.860], 0.805 [0.748,0.861], and 0.712 [0.621,0.804]), surpassing traditional clinical prognostic parameters (hazard ratio [HR]: 8.23, 95% CI: 5.43–12.47; p < 0.001). Using explainable AI (XAI) methods, we ensured model transparency and identified key TLS features-such as their location at the tumor periphery and their maturity state-as significant factors influencing prognosis and the efficacy of adjuvant therapy. The retrospective design without prospective validation and real-world clinical deployment is the main limitation of this study.

Together, these results highlight the potential utility of deep learning-based histopathological analysis for automated risk stratification in stage II CRC. In particular, our findings support the relevance of TLSs as a histological biomarker with potential implications for personalizing ACT decisions.

Approximately 20% of patients with stage II colorectal cancer (CRC) experience recurrence after surgery, but it remains challenging to identify which individuals are at high risk and might benefit from adjuvant chemotherapy.

Traditional clinicopathological markers provide some guidance but often lack sufficient accuracy for personalized treatment decisions.

There is a critical need for novel, scalable tools that can enhance risk stratification and inform therapeutic strategies using data routinely available in clinical practice.

We developed a deep learning model that automatically identifies key histological features to help predict patient risk and treatment benefit in stage II CRC.

The model identified tertiary lymphoid structure-like features as strong prognostic markers and integrated multiple histological views to improve prediction of disease recurrence.

The framework reliably stratified patients by relapse risk and identified subgroups that appeared to derive greater benefit from adjuvant chemotherapy.

This study demonstrates the potential of artificial intelligence to extract clinically meaningful features from standard histological slides, offering a non-invasive and cost-effective approach to guide treatment decisions.

Our findings support the future development of AI-assisted tools that can augment clinical decision-making with more consistent, interpretable, and personalized risk assessments in stage II CRC.

As a retrospective multicenter study, it is subject to selection bias and confounding, limiting causal inference.

In a retrospective analysis of whole slide images from individuals with stage II colorectal cancer, used to train a deep learning model, Zihan Zhao and colleagues identify histological features associated with risk of disease recurrence.