Constantly optimized mean teacher for semi-supervised 3D MRI image segmentation.

Area of Science:

• Medical Image Analysis
• Machine Learning
• Computer Vision

Background:

• Semi-supervised learning (SSL) methods, particularly the mean teacher model, show promise in magnetic resonance imaging (MRI) segmentation.
• Exponential moving average (EMA) in teacher models is hindered by unreliable unlabeled data, impacting prediction accuracy.
• Existing SSL approaches struggle with noise and data scarcity in medical imaging.

Purpose of the Study:

• To propose a novel framework for optimizing teacher models in SSL for medical image segmentation.
• To leverage expert-annotated data to improve teacher model reliability while retaining EMA benefits.
• To enhance robustness against noisy unlabeled data in MRI segmentation tasks.

Main Methods:

• A framework optimizing teacher models with reliable expert data, mitigating EMA's reliance on unlabeled images.
• Utilizing distinct data augmentation strategies: weak augmentation for the teacher and strong augmentation for the student model.
• Incorporating a double softmax mechanism to improve noise resistance and information learning.

Main Results:

• Achieved a Dice score of 91.02% on the Left Atrium (LA) dataset using only 20% labeled data, nearing supervised performance (91.14% with 100% data).
• Demonstrated significant improvements on the BraTS2019 dataset: 1.02% and 1.92% gains with 5% and 10% labeled data, respectively, over baseline methods.
• The proposed method shows competitive performance in semi-supervised medical image segmentation.

Conclusions:

• The proposed framework effectively optimizes teacher models in SSL for MRI segmentation, even with limited labeled data.
• The strategy of distinct data augmentation and double softmax enhances model robustness and performance.
• This approach presents a viable solution for medical image segmentation in resource-constrained, semi-supervised settings.