Student–teacher model based breast cancer classification approach with depthwise separable layers

Abstract

Breast cancer is among the most prevalent diseases affecting women worldwide and remains the leading cause of cancer-related mortality in women. Early and accurate diagnosis is critical for effective treatment and improved patient outcomes. Existing automated approaches, including Convolutional Neural Network (CNN)-AlexNet, CNN-Residual Network (ResNet), CNN-GoogleNet, Attention U-Net (AUNet), Multi-Task Learning Network (MTLNet), and Deep Supervision (DS) U-Net, face challenges such as limited data availability, overfitting, high computational requirements, and long training times. To address these limitations, a novel Residual connection assisted student–teacher distilled dual bidirectional vision transformer (Res-STdVT) is proposed for breast cancer detection and classification. The method begins with image acquisition from the Mini- Mammographic Image Analysis Society (MIAS) and CBIS-DDSM datasets, followed by effective pre-processing using an improved weighted guided filter (Imp-WeGF). The region of interest (ROI) is then extracted via upgraded fuzzy c-means clustering (Up-FCM). High-level and deep features are captured from the ROI using a Depthwise Separable Convolutional Inception Network (DSC-Inception Net), which are subsequently classified into benign and malignant categories by the proposed Res-STdVT model. Experimental results demonstrate that the model achieves superior performance, with accuracy, precision, recall, Intersection over Union (IOU) score, and dice score of 96.29 %, 98.88 %, 95.19 %, 98.91 %, and 98.23 % on Mini-MIAS, and 98.66 %, 96.25 %, 98.75 %, 96.43 %, and 98.84 % on Curated Breast Imaging Subset of the Digital Database for Screening Mammography (CBIS-DDSM), respectively. These findings suggest that the proposed method can assist radiologists and clinicians in reliable, automated breast cancer diagnosis, reducing workload and enhancing early detection.