Illuminating breast cancer malignancies: Lightweight histopathology computer vision classifier for precise breast cancer screening

Breast cancer, a major global health concern, underscores the pressing need for improved early detection methods to enhance patient outcomes. This study presents an innovative digital pathology classification approach that integrates Low-Rank Adaptation (LoRA) with the Vision Transformer (ViT) model. The proposed method enables more efficient breast cancer detection through a deep learning classifier that requires less training data. This approach outperforms traditional Convolutional Neural Network (CNN) models, including Residual Networks (ResNet), particularly in terms of performance and generalization capabilities in data-limited scenarios. Through extensive experiments and analyses across various dataset sizes, our streamlined classifier demonstrates higher training and testing accuracy (6 ​% better than AlexNet and 0.1 ​% better than ResNet-50 with over 99 ​% accuracy of our classifier) with less training parameters (59.85 million less than AlexNet and 24.85 million less than ResNet-50) in detecting breast cancer with full dataset. With less training data, our classifier performs 3 ​% better in testing accuracy than ResNet-50 with only 5 ​% dataset size. This work propels the advancement of sophisticated computer-aided diagnostic (CAD) systems, facilitating more rapid and accurate breast cancer detection, thereby significantly enhancing patient care outcomes. • Proposed model architecture achieves superior Breast benign tumors and malignant tumors classification accuracy over state-of-the-art methods on limited datasets. • Proposed lightweight model architecture with efficient LoRA tuning leverages ViT's visual representational power for rapid and accurate diagnosis.