ProsMAE: Multi-Source MAE Pretraining for ISUP Grade Classification

2026-07-09

AI Summary

Whole-slide images pose significant challenges for robust modeling due to their ultra-high resolution, staining variations, device heterogeneity, tissue artifacts, and scarce annotations. This work proposes ProsMAE, the first Masked Autoencoder pretraining framework tailored for multi-source histopathology data, enabling unsupervised representation learning across tissues and imaging conditions by jointly leveraging the PANDA, CAMELYON17, and BRACS datasets. The pretrained encoder is frozen and extended with a linear classification head (ProsCLS) for transfer to the prostate cancer ISUP grading task. Experimental results demonstrate that ProsMAE achieves significantly higher average validation quadratic weighted Kappa on independent PANDA splits compared to the standard MAE baseline, effectively enhancing model generalization and robustness.

Abstract

Whole slide images (WSIs) provide rich diagnostic information for computational pathology, but their gigapixel scale, stain variation, scanner differences, tissue artifacts, and limited expert annotation make robust model training challenging. This paper presents a multi-source Masked Autoencoder (MAE) framework, named ProsMAE, for histopathology representation learning. Tiles from Prostate cANcer graDe Assessment (PANDA), CAncer MEtastases in LYmph nOdes challeNge 2017 (CAMELYON17), and BReAst Carcinoma Subtyping (BRACS) are used for ProsMAE pretraining to expose the encoder to diverse tissue morphology and acquisition conditions. The learned encoder is transferred for International Society of Urological Pathology (ISUP) grade classification through ProsCLS, using a frozen encoder and a linear classification head. ProsMAE achieved a higher mean validation quadratic weighted kappa (QWK) than the vanilla MAE frozen linear-probe baseline under the evaluated disjoint PANDA split. Repeated-split evaluation remains necessary to further establish robustness across split compositions.

Problem

• whole slide images
• ISUP grade classification
• computational pathology
• stain variation
• limited annotation

Innovation

• Multi-source MAE
• Histopathology representation learning
• ISUP grade classification
• Self-supervised pretraining
• Cross-dataset generalization