USE-MiT: Attention-based model for breast ultrasound images segmentation

Early detection of breast cancer disease is crucial to enhancing patient outcomes through effective treatment. Ultrasound imaging, a simple, low-cost, and non-invasive technique, can help differentiate cystic from solid masses, mainly on the basis of the analysis of the detected anomalies’ boundaries. Automatic detection methods of mass boundaries in ultrasound images can reduce the dependence on the radiologist's experience for this analysis. We propose USE-MiT, a segmentation method for breast ultrasound images, based on a UNet architecture in which the encoder and decoder modules are interfaced through a configuration based on Squeeze and Excitation Attention modules, and the encoder structure is represented by a Mix Transformer. The model was trained and validated, with a 4-fold cross-validation, on the Breast Ultrasound Image Dataset, and was tested on the independent dataset, namely Breast-Lesions-USG. The experiments have demonstrated the efficiency of the model, achieving an overall Dice of 0.88 and an IoU of 0.64, outperforming the state-of-the-art. The source code is available at https://github.com/nbrancati/USE-MiT.