1Ph.D. Scholar, Computer Science and Engineering, Parul University, Vadodara, India
2Associate Professor, Computer Science and Engineering, Parul University, Vadodara, India
3Associate Professor, G.J. Patel Institute of Ayurvedic Studies and Research, CVM University, Vallabh Vidyanagar, Anand, India
Breast ultrasound imaging is commonly employed for the assessment of breast lesions, especially in women with dense breast tissue and in circumstances where mammography sensitivity is constrained. Nonetheless, interpreting ultrasound images continues to be difficult because of speckle noise, uneven backgrounds, and differences between operators. This study proposes a segmentation-guided, lesion-centric deep learning framework for the automated classification of breast ultrasound images. The framework combines U-Net-based lesion segmentation, region-of-interest (ROI) extraction, EfficientNetB0-based classification, and Gradient-weighted Class Activation Mapping (Grad-CAM) to make the model easier to understand. We tested the method on the Breast Ultrasound Images (BUSI) dataset, which had normal, benign, and malignant images. The training dataset went through preprocessing that included bilateral filtering, uniform resizing, and class-balanced augmentation. Before classification, segmentation outputs were used to find the areas of lesions. This reduced background noise and made it easier to learn features. The EfficientNetB0 classifier got an overall classification accuracy of 79.6% on the independent test set, with performance that was balanced across classes. Grad-CAM visualizations showed that the network's predictions were mostly based on lesion-relevant areas and not on artifacts in the surrounding tissue. The results show that segmentation-guided lesion-centric pipelines make it easier to understand and more reliable to diagnose breast cancer using ultrasound. They also lay the groundwork for computer-aided diagnostic systems that can be used in real life.
Keywords: Breast ultrasound, deep learning, U-Net segmentation, EfficientNet, explainable AI, medical image analysis
How to cite this article: Rana I, Shah J, Variya H. Segmentation-Guided Lesion-Centric Deep Learning Framework for Interpretable Breast Ultrasound Classification. Int J Drug Deliv Technol. 2026;16(12s): 323-327. DOI: 10.25258/ijddt.16.12s.35
Source of support: Nil.
Conflict of interest: None