HER2 amplification assessment in breast cancer in the HER2-SISH histopathological images using deep learning models

Human Epidermal Growth Factor Receptor 2 (HER2) status is a critical determinant of breast cancer prognosis and therapeutic strategy. Silver in situ Hybridization (SISH) enables HER2 gene assessment under bright-field microscopy while preserving tissue morphology, rendering it a clinically favoured alternative to fluorescence-based methods. Nevertheless, manual SISH slide interpretation is subjective, time-intensive, and susceptible to inter-observer variability, necessitating robust automated computational solutions. This work presents a systematic comparative evaluation of six deep learning architectures for multiclass HER2-SISH histopathology image classification across three diagnostic categories: Amplified, Non-Amplified, and ROI_Normal. The evaluated architectures span convolutional neural networks (MobileNetV3-Large, DenseNet121), transformer-based models (ViT-Small, DeiT-Small), and novel hybrid CNN-Transformer frameworks (DenseNet121-ViT, DenseNet121-DeiT). All models were trained under five-fold stratified cross-validation with a split-before-augmentation protocol and evaluated on an independent held-out test set over both the original and an augmented version of the publicly available HER2-SISH40x dataset. On the original dataset, DenseNet121 achieved the highest test accuracy of 82.10% with a macro F1-score of 77.33%. After offline rotational augmentation of the training partition, DenseNet121 achieved the best performance with test accuracy of 85.09% and macro F1-score of 80.61%, with ViT-Small (84.47%) and DenseNet121-ViT (83.23%) following closely. Per-class analysis revealed that ROI_Normal was consistently classified with the highest accuracy, whereas Amplified versus Non-Amplified discrimination remained the principal challenge across all evaluated models.