Self-Supervised Tumor Segmentation With Sim2Real Adaptation.

Area of Science:

• Medical Image Analysis
• Artificial Intelligence
• Computer Vision

Background:

• Accurate tumor segmentation is crucial for diagnosis and treatment planning.
• Unsupervised and self-supervised methods are needed to overcome the challenge of limited annotated medical data.

Purpose of the Study:

• To develop a novel self-supervised approach for tumor segmentation.
• To leverage synthetic data generation and a Sim2Real training regime for improved unsupervised segmentation.
• To achieve state-of-the-art performance on benchmark tumor segmentation datasets.

Main Methods:

• Proposed a 'layer-decomposition' proxy task for self-supervised pre-training.
• Developed a scalable pipeline for generating synthetic tumor data.
• Implemented a two-stage Sim2Real training regime: pre-training with simulated data, followed by self-training for adaptation.
• Evaluated on BraTS2018 (brain tumors) and LiTS2017 (liver tumors) datasets.

Main Results:

• Achieved state-of-the-art segmentation performance in unsupervised settings on BraTS2018 and LiTS2017.
• Outperformed existing self-supervised approaches in low-annotation tumor segmentation scenarios.
• Demonstrated that models trained on sufficiently randomized synthetic data generalize well to real tumor datasets.

Conclusions:

• The proposed self-supervised method effectively addresses the challenge of tumor segmentation with limited annotations.
• Synthetic data generation and the Sim2Real approach are viable strategies for unsupervised medical image segmentation.
• The 'layer-decomposition' proxy task shows promise for self-supervised learning in this domain.