An Efficient Semi-Supervised Framework with Multi-Task and Curriculum Learning for Medical Image Segmentation.

Area of Science:

• Medical Image Analysis
• Deep Learning
• Computer Vision

Background:

• Supervised deep learning for medical image segmentation faces challenges due to the scarcity and high cost of labeled data.
• Abundant unlabeled medical data is available in clinical settings, presenting an opportunity for improved model training.
• Existing semi-supervised methods often struggle to effectively leverage unlabeled data for enhanced segmentation performance.

Purpose of the Study:

• To develop a novel semi-supervised segmentation method that effectively utilizes unlabeled data.
• To improve generalization performance in medical image segmentation tasks with limited labeled data.
• To introduce a multi-task curriculum learning framework for enhanced segmentation.

Main Methods:

• A multi-task curriculum learning framework integrating a main segmentation task with two auxiliary tasks: feature regression and target detection.
• Auxiliary tasks generate pseudo-labels (image-level attributes and bounding boxes) to guide the main segmentation task.
• Incorporation of a bounding-box-based attention (BBA) module to address class imbalance and focus on target regions.
• Implementation of error tolerance mechanisms (inequality constraint, bounding-box amplification) to mitigate pseudo-label deviations.

Main Results:

• The proposed method demonstrates significantly improved segmentation performance on limited labeled datasets.
• Validation on ACDC2017 and PROMISE12 datasets shows superior results compared to fully supervised methods.
• Outperforms state-of-the-art semi-supervised segmentation techniques.
• The bounding-box-based attention module effectively enhances focus on relevant target regions.

Conclusions:

• The novel semi-supervised segmentation method via multi-task curriculum learning offers a powerful solution for medical image analysis.
• The approach effectively leverages unlabeled data to overcome limitations of small labeled datasets.
• The method shows great potential for clinical applications requiring accurate medical image segmentation.