Online Self-Distillation and Self-Modeling for 3D Brain Tumor Segmentation

Area of Science:

• Medical Image Analysis
• Artificial Intelligence in Medicine

Background:

• Supervised brain tumor segmentation faces challenges due to limited high-quality labeled data.
• Data privacy, cost, and ethical concerns restrict the availability of labeled medical datasets.

Purpose of the Study:

• To present a novel training framework, MOD (Masked Online Distillation), to improve supervised learning models for brain tumor segmentation with limited data.
• To enhance the efficacy of existing segmentation models without compromising computational efficiency during inference.

Main Methods:

• Integration of a plug-and-play component, MOD, into supervised learning frameworks.
• MOD utilizes an Online Tokenizer and Dense Predictor with self-distillation and self-modeling on masked patches.
• The component is excluded during inference to maintain computational efficiency.

Main Results:

• Models augmented with MOD demonstrated superior performance on 3D brain tumor segmentation tasks.
• Significant improvements in Dice coefficients and HD95 scores were observed on BraTS 2021 and MSD 2019 Task-01 Brain Tumor datasets.
• The approach achieved better results compared to leading 3D brain tumor segmentation baselines.

Conclusions:

• The proposed MOD framework effectively addresses the challenge of limited labeled data in brain tumor segmentation.
• MOD offers a computationally efficient method to boost the performance of supervised segmentation models.
• This approach holds promise for advancing automated medical image segmentation in resource-constrained scenarios.