HER-2/neu expression is observed in 15-20% of all breast cancers and is clinically relevant due to the several anti-HER2 treatment modalities. HER-2/neu status is assessed and quantified using immunohistochemistry (IHC). Although well-standardized in clinical practice, it is still subject to interobserver variability. Automating HER-2/neu scoring can improve accuracy, efficiency, consistency, and cost-effectiveness while reducing the pathologist's burden.
A total of 7,395 patch images were extracted automatically using a custom-built python application from 50 IHC-stained whole slide images (WSIs) from the HER2 challenge dataset of Warwick University. 1000×1000 pixels patches were created from the regions of interest of 20x zoomed WS image. All the patches were annotated to 0, 1+, 2+, and 3+ scores (i.e., classes) by two experts using an in-house built python application. Patches containing multiple classes or without any tumor tissues were not considered. Finally, a total of 6,488 patches (0: 1108, 1+: 276, 2+: 1442, 3+: 3662) were selected for five-fold subject-wise cross-validation after discarding 856 non-tumor patches and 51 multi-class patches. Three state-of-the-art deep learning models (GoogleNet, Densenet201, and Vit-base-r50) were validated to automatically classify the patches into four classes (0, 1+, 2+, and 3+).
Among the three investigated models, the Vit-base-r50 model showed the best performance with overall accuracy, precision, and sensitivity of 90.02%, 90.11% and 90.09%, respectively. It achieved a sensitivity of 97.64%, 75.07%, 84.19% and 85.75%, while the second-best model (Densenet201) achieved a sensitivity of 95.51%, 75.79%, 78.54% and 76.86% for 0, 1+, 2+, and 3+ classes, respectively.
The proposed deep learning-based approach for automatic detection and assessment of HER-2/neu expression in breast cancer shows promising results. This procedure can offer a cost and time-effective alternative to producing clinically relevant results and may help pathologists in a proper assessment of HER-2/neu expression.
Dr. Semir Vranic.
Has not received any funding.
All authors have declared no conflicts of interest.