MTANS: Multi-Scale Mean Teacher Combined Adversarial Network with Shape-Aware Embedding for Semi-Supervised Brain...

Area of Science:

• Medical Image Analysis
• Artificial Intelligence
• Neuroscience

Background:

• Accurate brain lesion segmentation is crucial for diagnosing and treating various brain diseases.
• Deep learning methods require extensive high-quality annotations, which are time-consuming to obtain.
• Semi-supervised learning offers a solution by utilizing limited labeled data alongside unlabeled data.

Purpose of the Study:

• To develop a novel semi-supervised segmentation framework for brain lesion images.
• To leverage both labeled and unlabeled data to enhance segmentation performance.
• To reduce the manual effort required for image annotation in clinical settings.

Main Methods:

• A framework combining an improved mean teacher and adversarial network for segmentation.
• Utilizing student and teacher models for segmentation and signed distance map generation.
• Employing a discriminator network for feature extraction and signed distance map distinction.
• Incorporating multi-scale feature consistency loss and a shape-aware embedding scheme.

Main Results:

• The proposed method effectively utilizes unlabeled data for improved brain lesion segmentation.
• Outperformed supervised baselines and other state-of-the-art semi-supervised methods on multiple public datasets (ISBI 2015, ISLES 2015, BRATS 2018).
• Demonstrated strong performance in segmenting multiple sclerosis lesions, ischemic stroke lesions, and brain tumors.

Conclusions:

• The novel semi-supervised framework significantly enhances brain lesion segmentation accuracy.
• The method effectively reduces the need for extensive manual annotation, easing clinical workload.
• This approach is suitable for joint training with limited labeled and abundant unlabeled data, advancing medical image analysis.