Which Attention Helps? A Slide-Level Ablation of Attention-Augmented CNNs for Serous Ovarian Carcinoma Grading

Pretrained MobileNet with a differential-attention module best distinguishes high-grade from low-grade serous carcinoma on the UBC-OCEAN slide-level task, achieving AUC 0.862 ± 0.005.

• Why it matters: Slide-level leakage can inflate reported accuracy for HGSC versus LGSC grading, leaving uncertainty about which attention designs truly help under rigorous evaluation.
• What they did: Using 259 whole-slide images (217 HGSC, 42 LGSC) with five-fold slide-grouped cross-validation, stain normalization, slide-level pooling, and attention ablations, models were compared to Phikon with DeLong/McNemar tests.
• The result: Pretraining significantly outperformed Squeeze-and-Excitation (p = 0.002) and the no-attention baseline (p = 0.019), while adding spatial attention to channel attention showed no detectable gain (0.824 to 0.822; p = 0.50).

AI summary of the abstract · may contain errors · not medical advice · read the paper.