Investigating the Impact of Histopathological Foundation Models on Regressive Prediction of Homologous Recombination Deficiency

TL;DR

This work investigates whether histopathology foundation models can improve regressive biomarker prediction by forecasting a continuous $HRD$ score from WSIs using MIL. Through careful comparison of five foundation models and MIL aggregators (attMIL and SuRe Transformer) across TCGA cohorts and CPTAC external datasets, the authors demonstrate consistent performance gains over a RetCCL baseline, with UNI-2 and Virchow-2 often leading. They address data imbalance with a distribution-based upsampling strategy that boosts recall and balanced accuracy for underrepresented HRD+ cases, while ablation studies highlight the relative importance of instance selection and sampling strategies over sheer bag size. The results underscore the potential of large-scale, histopathology-ready pretraining to yield transferable, regression-ready representations for precision oncology and other regressive biomarker tasks.

Abstract

Foundation models pretrained on large-scale histopathology data have found great success in various fields of computational pathology, but their impact on regressive biomarker prediction remains underexplored. In this work, we systematically evaluate histopathological foundation models for regression-based tasks, demonstrated through the prediction of homologous recombination deficiency (HRD) score - a critical biomarker for personalized cancer treatment. Within multiple instance learning frameworks, we extract patch-level features from whole slide images (WSI) using five state-of-the-art foundation models, and evaluate their impact compared to contrastive learning-based features. Models are trained to predict continuous HRD scores based on these extracted features across breast, endometrial, and lung cancer cohorts from two public medical data collections. Extensive experiments demonstrate that models trained on foundation model features consistently outperform the baseline in terms of predictive accuracy and generalization capabilities while exhibiting systematic differences among the foundation models. Additionally, we propose a distribution-based upsampling strategy to mitigate target imbalance in these datasets, significantly improving the recall and balanced accuracy for underrepresented but clinically important patient populations. Furthermore, we investigate the impact of different sampling strategies and instance bagsizes by ablation studies. Our results highlight the benefits of large-scale histopathological pretraining for more precise and transferable regressive biomarker prediction, showcasing its potential to advance AI-driven precision oncology.

Figures (7)

• Figure 1: The overall architecture of our workflow. WSIs are tessellated and normalized before patches are encoded via feature extraction models to create patient-specific bags of features. From these, $n$ instances of dimension $d$ are sampled (bagsize $n$ x embedding dimension $d$), and processed through one of two aggregation architectures: an attention-based MIL (attMIL) or a SuRe Transformer. The final slide-level representation is passed to a regression head to predict the continuous HRD score ($y_B$).
• Figure 2: Performance rank of TCGA internal crossvalidation. The rank is computed by comparing the rank of each model on each individual fold.
• Figure 3: Performance rank for external validation of the same cancer type. The rank is computed by comparing the rank of each model on each individual fold.
• Figure 4: Original vs. upsampled distribution of the TCGA LUAD after applying the distribution-based upsampling algorithm
• Figure 5: Binned RMSE of attMIL architecture trained with UNI feature extractor with (orange) and without (blue) upsampling the training data