Attention enhanced DeepLabV3+ framework for breast cancer classification

Timely detection of breast abnormalities with medical imaging is essential for providing more effective treatment, yet remains challenging due to subtle lesion characteristics, low contrast, and complex background structures. Many existing deep learning–based classification approaches rely on whole-image analysis, which may result in missed lesions and more chances of false-negative predictions. To overcome these limitations, this study proposes ADSCNet framework, a novel two stage segmentation-guided classification approach designed to enhance diagnostic reliability in breast cancer classification. In stage 1, for precise lesion segmentation DeepLabV3+ with ResNet50 encoder integrated with CBAM into the decoder, where high and low level features are fused is used. Comparative evaluation is performed using several advanced semantic segmentation architectures: U-Net, attention-enhanced U-Net++, and standard DeepLabV3+. The comparison demonstrates that DeepLabV3+ with CBAM outperforms with Mean Dice of 90.14%, Median Dice of 97.19%, and Mean IoU of 86.3% by effectively emphasizing salient lesion regions through its enhanced attention mechanism. The generated segmented masks are then guided to stage 2 for classification where, a lightweight and computationally efficient deep learning model, modified DenseNet121 integrated with CBAM after each of four dense block is employed to classify lesions into cancerous, non-cancerous and normal cases. The segmentation-guided inputs enables the classifier to concentrate on diagnostically meaningful regions, resulting in improved sensitivity and a noticeable reduction in false-negative cases, which is particularly critical in clinical screening applications. In the classification stage model achieves F1-scores of 0.9945, 0.9873, and 0.9947 for normal, benign, and malignant cases respectively, with a good recall for malignant cases. Additionally, an ablation study is also conducted to validate the contribution of changes made in the framework. Further to enhance the generalization capability of proposed ADSCNet framework, cross data validation is performed using UDIAT dataset. Experimental results confirm that the proposed framework achieves improved classification performance. By combining precise lesion segmentation, lightweight classification, and cross data validation, this work demonstrates a segmentation guided approach for breast cancer classification, that may be explored further for successful integration in computer-aided diagnostic systems.