Deep learning assessment of metastatic relapse risk from digitized breast cancer histological slides

Abstract

Accurate risk stratification is critical for guiding treatment decisions in early breast cancer. We present an artificial intelligence (AI)-based tool that analyzes digitized tumor slides to predict 5-year metastasis-free survival (MFS) in patients with estrogen receptor-positive, HER2-negative (ER + /HER2 - ) early breast cancer (EBC). Our deep learning model, RlapsRisk BC, independently predicts MFS and provides significant prognostic value beyond traditional clinico-pathological variables (C-index 0.81 vs 0.76, p < 0.05). Applying a 5% MFS event probability threshold stratifies patients into low- and high-risk groups. After dichotomization, combining RlapsRisk BC with clinico-pathological factors increases cumulative sensitivity (0.69 vs 0.63) and dynamic specificity (0.80 vs 0.76) compared to clinical factors alone. Expert analysis of high-impact regions identified by the model highlights well-established morphological features, supporting its interpretability and biological relevance.

Fig. 1. RlapsRisk BC algorithm overview.

The left panel illustrates the algorithm’s processing steps, while the right panel details the training procedure and the overall model architecture designed to predict the risk group.

Fig. 2. Forest plot of the adjusted RlapsRisk BC score HRs on the prediction of 5-year metastasis-free survival on the CANTO cohort.

Each square of the forest plot represents the HR of the RlapsRisk BC score (a continuous variable) adjusted for the prognostic clinico-pathological factors in the subgroup of patients defined by the variable category in the first column of the table. The 95% HR confidence intervals, computed using the Wald method, are represented by the horizontal lines. Histological tumor grade 1 and low clinical risk group were removed as no MFS events were recorded in these subgroups. Hazard ratios (HRs) and 95% confidence intervals were estimated using a Cox proportional hazards model. p Values correspond to two-sided Wald tests evaluating the null hypothesis HR = 1 and were adjusted for multiple comparisons using the Benjamini–Hochberg procedure. *p < 0.1, **p < 0.05, ***p < 0.01.

Fig. 3. Metastases-free survival of patients stratified according to RlapsRisk BC classifier (left), Clinical score classifier (middle), and RR Combined classifier (right) among patients from the CANTO validation cohort.

p Values were computed using the two-sided log-rank test to compare survival distributions between groups. Shaded areas represent 95% confidence intervals estimated using the Greenwood formula.

Fig. 4. Metastases-free survival of patients stratified according to RlapsRisk BC classifier (left), Clinical score classifier (center), and RR Combined classifier (right) among patients from the CANTO intermediate clinical risk validation cohort.

p Values were computed using the two-sided log-rank test to compare survival distributions between groups. Shaded areas represent 95% confidence intervals estimated using the Greenwood formula.

Fig. 5. Interpretability of the RlapsRisk BC model.

The methodology applied to generate the data used for this analysis is outlined. To streamline pathologist review, slides were sampled to express as much diversity as possible in the high, low, and intermediary predicted risks. Shapley values were computed to evaluate the positive or negative contributions of clusters of similar tiles. A selection of the most contributing clusters of tiles was made, and random clusters were also sampled from the distribution of Shapley values. An expert pathologist annotated the data, blinded to any information about contribution or risk.

Fig. 6. A selection of the histological features annotated by an expert pathologist.

The proportion of tiles annotated as containing a particular histological feature within clusters of tiles linked to high and low risk, as well as randomly selected ones, is presented alongside the significance of p values from the Chi-square test of independence computed between each group, adjusted for multiple testing using the Benjamini–Hochberg procedure. For each histological feature, the region of a WSI containing the annotated cluster of tiles is displayed, extracted both from the CANTO H&E dataset (upper histological image) and from the CANTO HES dataset (lower histological image). ns p > 0.05, **p < 0.01, ***p < 0.001, ****p < 0.0001.