A multi-stage training and deep supervision based segmentation approach for 3D abdominal multi-organ segmentation

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Radiology

Background:

• Accurate segmentation of abdominal organs in 3D CT images is crucial for medical diagnosis, cancer treatment planning, and radiotherapy.
• Current deep learning models struggle with complex organ distributions, limited labeled data, and diverse organ structures, leading to suboptimal segmentation accuracy.

Purpose of the Study:

• To develop an improved deep learning approach for 3D abdominal multi-organ segmentation.
• To address challenges of data scarcity, model convergence, and segmentation accuracy in abdominal CT image analysis.

Main Methods:

• A novel multi-stage training strategy combined with a deep supervision model incorporating an attention mechanism (DLAU-Net).
• Integration of a pseudo-labeling technique to overcome limitations of scarce labeled data.
• Utilized a large dataset from the FLARE 2023 Challenge for validation.

Main Results:

• The proposed DLAU-Net achieved an average organ accuracy (AVG) of 90.5% and a Dice Similarity Coefficient (DSC) of 89.05%.
• Demonstrated superior performance in training speed, handling data diversity, and segmenting critical organs like the liver, spleen, and kidneys.
• Significantly outperformed existing comparative methods in segmentation tasks.

Conclusions:

• The developed multi-stage training and DLAU-Net approach effectively enhances 3D abdominal multi-organ segmentation accuracy and efficiency.
• The method shows strong generalizability and addresses key challenges in medical image analysis, offering a promising tool for clinical applications.