67P The utility of a deep learning-based approach in HER-2/neu assessment in breast cancer

HER-2/neu expression is observed in 15-20% of all breast cancers and clinically relevant due to the several anti-HER2 treatment modalities. status assessed quantified using immunohistochemistry (IHC). Although well-standardized clinical practice, it still subject interobserver variability. Automating scoring can improve accuracy, efficiency, consistency, cost-effectiveness while reducing pathologist's burden. A total 7,395 patch images were extracted automatically a custom-built python application from 50 IHC-stained whole slide (WSIs) HER2 challenge dataset Warwick University. 1000×1000 pixels patches created regions interest 20x zoomed WS image. All annotated 0, 1+, 2+, 3+ scores (i.e., classes) by two experts an in-house built application. Patches containing multiple classes or without any tumor tissues not considered. Finally, 6,488 (0: 1108, 1+: 276, 2+: 1442, 3+: 3662) selected for five-fold subject-wise cross-validation after discarding 856 non-tumor 51 multi-class patches. Three state-of-the-art deep learning models (GoogleNet, Densenet201, Vit-base-r50) validated classify into four (0, 3+). Among three investigated models, Vit-base-r50 model showed best performance with overall precision, sensitivity 90.02%, 90.11% 90.09%, respectively. It achieved 97.64%, 75.07%, 84.19% 85.75%, second-best (Densenet201) 95.51%, 75.79%, 78.54% 76.86% classes, The proposed learning-based approach automatic detection assessment cancer shows promising results. This procedure offer cost time-effective alternative producing results may help pathologists proper expression.