Semi-supervised abdominal multi-organ segmentation by object-redrawing.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Computer Vision

Background:

• Multi-organ segmentation is crucial for medical imaging applications like radiation therapy planning and quantitative analysis.
• Manual segmentation is time-consuming and lacks reproducibility; deep learning methods often require extensive labeled data.
• Existing semi-supervised learning (SSL) methods for abdominal multi-organ segmentation have limitations.

Purpose of the Study:

• To introduce a novel SSL approach for abdominal multi-organ segmentation that leverages unlabeled data.
• To enhance the performance of deep neural networks in segmenting abdominal organs.
• To incorporate a redrawing network to correct segmentation errors and improve accuracy.

Main Methods:

• A novel SSL method using three interconnected neural networks: segmentation, teacher, and redrawing networks.
• The segmentation network is trained using labeled and unlabeled data with consistency learning (Mean-Teacher model).
• A redrawing network generates corrected images from CT scans, preserving anatomical information, to reduce segmentation errors during a readjustment phase.

Main Results:

• The proposed SSL method was evaluated on the BTCV and AMOS datasets.
• It consistently outperformed state-of-the-art SSL methods (MT, DTC) and supervised learning approaches.
• Superior segmentation performance for abdominal organs was achieved, demonstrating effectiveness with limited labeled data.

Conclusions:

• The novel SSL approach effectively addresses challenges in abdominal multi-organ segmentation.
• Integrating a redrawing network and utilizing unlabeled data significantly improves segmentation accuracy.
• This method shows promise for enhancing precision and efficiency in medical imaging segmentation.