Cascaded Self-Supervision to Advance Cardiac MRI Segmentation in Low-Data Regimes

Area of Science:

• Medical Image Analysis
• Deep Learning
• Cardiovascular Imaging

Background:

• Supervised deep learning for medical image segmentation requires extensive labeled data, which is costly and time-consuming to acquire, especially for pixel-level 3D annotations.
• Limited labeled data restricts model performance in medical image segmentation tasks.
• Unlabeled data is often abundant and can be leveraged through semi- or self-supervised learning (SSL) to enhance model training.

Purpose of the Study:

• To investigate and compare popular SSL strategies, including Transformation Consistency, Student-Teacher, and Pseudo-Labeling, for cardiac MRI segmentation.
• To evaluate the effectiveness of these SSL methods, individually and in combination, under limited labeled data scenarios.
• To propose and validate a novel cascaded Self-Supervision methodology combining Pseudo-Labeling and a cascaded Student-Teacher model.

Main Methods:

• Evaluation of Transformation Consistency, Student-Teacher, and Pseudo-Labeling SSL strategies on 2D and 3D cardiac MRI datasets (ACDC, MMWHS).
• Assessment of performance across various low-data regimes with decreasing amounts of labeled training data.
• Implementation of a cascaded Self-Supervision approach integrating Pseudo-Labeling and a self-supervised cascaded Student-Teacher model.

Main Results:

• All investigated SSL methods outperformed the supervised baseline and existing state-of-the-art self-supervised approaches in all scenarios.
• The proposed cascaded Self-Supervision method achieved significant Dice Similarity Coefficient (DSC) improvements of 10.17% (ACDC) and 6.72% (MMWHS) over the low-data supervised approach in the very-low-labeled data regime.
• The proposed method substantially reduced the performance gap compared to fully supervised models, demonstrating the efficacy of incorporating unlabeled data.

Conclusions:

• Cascaded Self-Supervision, combining Pseudo-Labeling and a cascaded Student-Teacher model, is the most effective methodology for cardiac MRI segmentation.
• Leveraging unlabeled data through SSL is consistently beneficial for improving segmentation accuracy when labeled data is scarce.
• The findings highlight the potential of SSL to overcome data limitations in medical image analysis and enhance clinical applications.