A comparison between Deep Learning architectures for the assessment of breast tumor segmentation using VSI ultrasound protocol

Automatic breast tumor ultrasound segmentation is one of the most critical components in the development of tools for breast cancer diagnosis. Several deep learning algorithms have been tested with public and private datasets but none of them has been designed for asynchronous protocol ultrasound acquisition. In this work, a dataset collected through the Volume Sweep Imaging protocol for breast ultrasound (VSI-B) was used. A comparative analysis of convolutional neural networks for segmentation was carried out, including the preliminary stages of data cleaning and preprocessing. The networks evaluated were: U-NET, Attention U-NET, Residual U-NET, and multi-input attention U-NET; among which the multi-input attention U-NET was identified as the best model, achieving a 72.45% Dice coefficient after a leave-one-out cross-validation with 53 patients. The results show that these semantic segmentation approaches could be useful for automatic tumor segmentation, particularly for asynchronous acquisitions such as VSI-B.