BCViT++: Locality-Aware Large-Kernel Vision Transformer for Breast Cancer Histopathology Classification

Automated categorization of breast histopathology pictures can benefit computational pathology, however tissue appearance changes with optical magnification, and annotated datasets remain tiny for Vision Transformers (ViTs). In this study, we propose BCViT++, a magnification-specific framework for binary categorization of benign and malignant breast tumor pictures. This pipeline incorporates channel-wise principle component analysis (PCA) reconstruction with a variance threshold of at least 99%, training-set class balancing, geometric augmentation, Locality Self-Attention (LSA), and Large Kernel Attention (LKA). LSA masks token self-attention with a diagonal matrix and learns the attention temperature, while LKA refines value representations using depth-wise convolutions with a larger effective receptive field. All BreakHis experiments were done separately at 40×, 100×, 200× and 400×. We obtained an accuracy of 86.57%, 86.87%, 90.37% and 90.44% using a PCA-ViT baseline, respectively. The whole BCViT++ framework improved these accuracies to 93.53%, 91.17%, 96.79%, and 96.45% with absolute improvements of 6.96, 4.30, 6.42 and 6.01 percentage points, respectively. The best performance was achieved at 200× with accuracy of 98.55%, sensitivity of 96.79%, specificity of 96.80%, F1-score of 97.66% and MCC of 0.926. These results show consistent overall improvements over the PCA-ViT baseline for the studied image-level technique.