Multi-view stereoscopic attention network for 3D tumor classification in automated breast ultrasound

As the new generation of breast cancer screening tools, automated breast ultrasound (ABUS) has some difficulties of the long interpretation time in clinical application. A computer-aided diagnosis algorithm can assist diagnosticians in locating and classifying benign and malignant tumors in ABUS images quickly. However, locating and classifying ABUS images is challenging because they have inapparent and small lesion regions and blurry discriminative regions. This paper proposes the multi-view stereoscopic attention network (MVSA-Net) to address the challenges. First, MVSA-Net crops the three-dimensional (3D) lesion regions through the 3D localization unit. Second, MVSA-Net classifies the cropped lesion regions through the classification unit. The localization unit can pay more attention to lesion regions by constructing the stereoscopic attention module. Besides, the localization unit constructs the 3D localization tensor through the split output design. The classification unit can extract the stereoscopic features through the stereoscopic view and the plane features through the plane view based on the Transformer network. It also fuses the two features through the level connection design. Besides, it can more focus on the borderline region of the tumor based on the region-prior module. The paper uses 172 ABUS images to train MVSA-Net, and 42 ABUS images to test MVSA-Net. On the test set, the accuracy and AUC of MVSA-Net are 95.24% and 99.76%, respectively. MVSA-Net can assist diagnosticians in decreasing the interpretation time of ABUS images, thus helping more people screen breast tumors earlier.

Automated summary

This paper proposes a multi-view stereoscopic attention network (MVSA-Net) to address the challenges of locating and classifying lesions in automated breast ultrasound (ABUS) images. MVSA-Net crops the three-dimensional lesion regions and classifies them through the stereoscopic features extracted from the Transformer network. The results show that MVSA-Net achieves high accuracy and AUC, indicating its potential to assist diagnosticians in reducing the interpretation time of ABUS images.