Attention-Guided Deep Learning Framework for Ovarian Cancer Subtype Classification

Ovarian cancer histotype classification is challenging because of substantial morphological heterogeneity and subtle subtype-specific features. This work presents a controlled evaluation of high-resolution pathology representations and slide-level aggregation methods for automated classification of five ovarian cancer subtypes from whole-slide histopathology images. Precomputed CONCH patch embeddings were aggregated using mean pooling, max pooling, gated attention-based multiple-instance learning, and a max-pooling cascade. The models were evaluated on 513 non-TMA whole-slide images from UBC-OCEAN using leakage-controlled five-fold cross-validation, with each slide receiving exactly one out-of-fold prediction. Gated ABMIL achieved a balanced accuracy of 81.39% and a macro-F1 score of 81.96%. Max pooling produced the highest numerical performance, with a balanced accuracy of 81.44%, macro-F1 of 82.59% (95% CI: 78.84-86.26%), macro-AUROC of 96.99%, and macro-AUPRC of 90.79%. However, paired slide-level bootstrap comparisons found no statistically significant differences among the CONCH aggregation strategies after Holm correction. Compared with the EfficientNet-B0 thumbnail baseline, CONCH max pooling improved macro-F1 by 30.30 percentage points and balanced accuracy by 25.70 percentage points, with both paired bootstrap confidence intervals excluding zero. Attention weights enabled visualization of influential patches, although these regions were not independently validated by pathologists. The findings show that high-resolution pathology foundation-model representations support ovarian cancer subtyping, while greater aggregation complexity does not necessarily improve performance. External multi-institutional validation is required before clinical generalizability can be established.