Benchmarking Histopathology Foundation Models for Ovarian Cancer Bevacizumab Treatment Response Prediction from Whole Slide Images

TL;DR

This work tackles the challenge of predicting bevacizumab treatment response in ovarian cancer using histopathology WSIs. It benchmarks a suite of histopathology foundation encoders across multiple MIL strategies, formulating the task with binary response via $L_{CE}$ and survival through the Cox partial likelihood $L_{COX}$ while keeping encoders frozen. The study demonstrates that domain-specific, self-supervised histopathology encoders outperform natural-image pretraining, achieving an accuracy of 72.5% and an AUC up to 0.86 for response prediction, and a c-index around 0.64 for survival with significant risk stratification. Furthermore, the analysis highlights high-attention tumor regions as potentially prognostic imaging biomarkers and points to transformer-based architectures as advantageous for histopathology tasks, underscoring the potential of imaging-based biomarkers to guide personalized bevacizumab therapy.

Abstract

Bevacizumab is a widely studied targeted therapeutic drug used in conjunction with standard chemotherapy for the treatment of recurrent ovarian cancer. While its administration has shown to increase the progression-free survival (PFS) in patients with advanced stage ovarian cancer, the lack of identifiable biomarkers for predicting patient response has been a major roadblock in its effective adoption towards personalized medicine. In this work, we leverage the latest histopathology foundation models trained on large-scale whole slide image (WSI) datasets to extract ovarian tumor tissue features for predicting bevacizumab response from WSIs. Our extensive experiments across a combination of different histopathology foundation models and multiple instance learning (MIL) strategies demonstrate capability of these large models in predicting bevacizumab response in ovarian cancer patients with the best models achieving an AUC score of 0.86 and an accuracy score of 72.5%. Furthermore, our survival models are able to stratify high- and low-risk cases with statistical significance (p < 0.05) even among the patients with the aggressive subtype of high-grade serous ovarian carcinoma. This work highlights the utility of histopathology foundation models for the task of ovarian bevacizumab response prediction from WSIs. The high-attention regions of the WSIs highlighted by these models not only aid the model explainability but also serve as promising imaging biomarkers for treatment prognosis.

Figures (5)

• Figure 1: a) Patient distribution across the different subtypes in the ovarian bevacizumab response dataset. b) Slide distribution across the different subtypes showing the label distribution for each subtype. The ovarian cancer subtypes include clear cell carcinoma (CC), endometrioid carcinoma (EC), mucinous carcinoma (MC), peritoneal serous papillary carcinoma (PsPC), papillary serous carcinoma (PsC), unclassifed carcinoma (UC). The slide labels include "Effective" corresponding to the patients with a favorable response to bevacizumab treatment and "Not effective" corresponding the non-responders of the bevacizumab treatment.
• Figure 2: Whole slide image (WSI) analysis pipeline depicting the steps involved in the processing of WSIs that includes pre-processing (tissue area masking, patch sampling, color normalization), patch-level feature extraction from foundation models, aggregating patch-level features to produce a slide-level representation using MIL model, and prediction of treatment response. Note that we only train the MIL model along with the MLP layers while training our models with the cross-entropy loss ($L_{CE}$) for a binary classification task of treatment effectiveness prediction or the Cox partial likelihood loss ($L_{COX}$) for the time-to-event regression task of survival prediction.
• Figure 3: a) Receiver Operating Characteristic (ROC) curves showing the treatment response prediction performance along with the AUC score for different encoders. b) Top 5 high-attention patches sorted in the descending order of the attention values (left to right) for the best-performing models from our experiments namely, 1) CTransPath with Clam-SB, 2) CTransPath with VarMIL, 3) Lunit-Dino with ABMIL.
• Figure 4: Kaplan-Meier plots stratifying the high- and low-risk cases in the test set across the 5 best performing models from Table \ref{['table_survival_results']}. a) CTransPath with ABMIL b) CTransPath with VarMIL c) Swin with TransMIL d) UNI with Clam-SB e) UNI with Clam-MB.
• Figure 5: Kaplan-Meier plots stratifying the high- and low-risk serous cases in the test set across the 5 best performing models from Table \ref{['table_survival_results']}. a) CTransPath with ABMIL b) CTransPath with VarMIL c) Swin with TransMIL d) UNI with Clam-SB e) UNI with Clam-MB.