Segmentation of Intra-operative Ultrasound Using Self-supervised Learning Based 3D-ResUnet Model with Deep Supervision

Abstract

Intra-operative ultrasound (iUS) is a robust and relatively inexpensive technique to track intra-operative tissue shift and surgical tools. Automatic algorithms for brain tissue segmentation in iUS, especially brain tumors and resection cavity can greatly facilitate the robustness and accuracy of brain shift correction through image registration, and allow easy interpretation of the iUS. This has the potential to improve surgical outcomes and patient survival rates. In this paper, we have proposed a self-supervised two-stage model for the Intra-operative ultrasound (iUS) task. In the first stage, we trained the encoder of our proposed 3DResUNet model using the self-supervised contrastive learning. The self-supervised learning offers the promise of utilizing unlabeled data. The training samples are used in self-supervision to train the encoder of the proposed 3DResUNet model and utilized this encoder as a pre-trained weight for the Intra-operative ultrasound (iUS) segmentation. In the second stage, the pre-trained weighted-based 3DResUNet proposed model was used to train on the training dataset for iUS segmentation. Experiment on CuRIOUS -22 challenge showed that our proposed solution showed significantly better performance before, during, and after Intra-operative ultrasound (iUS) segmentation. The code is publicly available (https://github.com/RespectKnowledge/SSResUNet_Intra-operative-ultrasound-iUS-Tumor-Segmentation).