Transformers meet CNNs for insights into breast mass classification from histopathological images

Breast cancer remains one of the leading causes of cancer-related deaths among women worldwide, highlighting the critical need for accurate histopathological diagnosis and reliable decision-support systems to improve diagnostic sensitivity and reduce false-negative outcomes.

In this research, a deep learning-based approach for binary classification of breast cancer into benign and malignant categories utilizing histopathological images is presented. A dataset comprising 10,000 high-resolution histopathology images was used to evaluate the execution of two vision models: Swin Transformer V2 and ConvNeXt V2. Swin Transformer V2, a progressive vision transformer with shifted window self-attention, and ConvNeXt V2, a modern convolutional neural network motivated by transformer plans, were fine-tuned and tested for their adequacy in feature representation and classification accuracy.

The experimental results demonstrate that Swin Transformer V2 consistently outperforms ConvNeXt V2 across all evaluation metrics, achieving a peak classification accuracy of 0.985, which reflects its superior capability in capturing subtle morphological and contextual variations in histopathological tissues.

The attention-driven global feature modeling in Swin Transformer V2 enables more discriminative representations compared to convolutional inductive biases, particularly for complex cellular patterns. These findings suggest that transformer-based architectures offer significant advantages over modern CNNs for histopathological breast cancer classification, and they hold substantial potential for advancing computer-aided diagnosis systems in digital pathology. The comparative insights provided in this study can guide the selection of robust deep learning models for scalable and reliable clinical decision-support systems.