Self-supervised learning-leveraged boosting ultrasound image segmentation via mask reconstruction

Deep neural networks have improved universal machine learning tasks due to their excellent ability to learn rich semantic features from high-dimensional data. However, large manually labeled datasets are usually required for machine learning. These datasets are often expensive and impractical to obtain due to the time required by professionals to label the data and data protection issues. Self-supervised pretraining can be used to initialize the model, which can then be fine-tuned on a small amount of labeled data. In this study, we aimed to develop a self-supervised learning algorithm that could be used to segment ultrasound images automatically. To achieve this aim, we proposed an pretext task of mask reconstruction to extract relevant semantic features from unlabeled data. Additionally, we also present a novel segmentation network named SEAT-U-Net for further improve the segmentation ability of the model. This network utilizes the Transformer Encoder and U-Net Encoder to extract and fuse relevant features through channel attention mechanisms. The training and validation of the algorithm were performed on breast and thyroid datasets. Our algorithm achieved a higher segmentation accuracy with less labeled data for thyroid datasets. When we apply the model trained in the thyroid dataset to the breast dataset, it also achieves very good result, which validates the effectiveness and robustness of our SEAT-U-Net.