Attention-Guided EfficientNet-B3 with Supervised Tumor Segmentation for Breast Ultrasound Classification
Abstract
Breast ultrasound interpretability is complicated because of issues such as speckle noise, low contrast, different sizes and shapes of lesions, and unclear edges. Many automated decision systems fail to demonstrate the accuracy of their localization process due to single single-time testing. This work proposes a two-branched deep learning framework for three-class breast ultrasound classification and tumor segmentation. The constructed model consists of an ImageNet-pretrained-efficient EfficientNet-B3 as an encoder with a channel attention module for the classification part and a U-Net architecture for pixel segmentation of the images. The output maps for Grad-CAM reproducibilities are separated from the results of educated segmentation. After application of the framework in five corresponding trials on the BUSI dataset composed of 780 images, it demonstrated the following classification results: accuracy - 87.69% ± 2.74%, Macro-F1 - 86.64% ± 2.95%, Macro-AUC - 95.67% ± 1.26%. For segmentation purposes, the framework achieved following metrics: Overlap scores: Dice = $\mathbf{7 0. 3 3 2} \boldsymbol{\%} \pm \mathbf{4. 6 1} \boldsymbol{\%}, \mathbf{I o U} \boldsymbol{=} \mathbf{6 2. 2 9 \%} \pm \mathbf{4. 9 6 \%}$ and specificity $\boldsymbol{=} \mathbf{9 8. 2 5 \%}$ ± 0.42%. Achieving confirmation of its performance, the proposed framework was evaluated in comparison to EfficientNet-B3, ResNet-50, and Swin-Tiny algorithms.