Body Part Regression With Self-Supervision.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Radiology

Background:

• Body part regression uses self-supervised learning for content navigation in computed tomography (CT) scans.
• Defining a unified coordinate system for CT scans is difficult due to variations in resolution, contrast, sequences, and anatomy.
• Supervised learning methods are not easily applicable to this challenge.

Purpose of the Study:

• To propose an annotation-free method, the blind-unsupervised-supervision network (BUSN), for body part regression in CT scans.
• To address the limitations of existing methods in handling diverse CT scan data.
• To improve the accuracy and consistency of spatial localization in medical imaging.

Main Methods:

• Developed BUSN using 1030 multi-center CT scans without manual annotation.
• Implemented a self-correction mechanism where BUSN refines unsupervised learning predictions for new supervision.
• Integrated a neighbor message passing (NMP) scheme for enhanced prediction consistency.
• Introduced a pre-processing pipeline incorporating BUSN for 3D multi-organ segmentation.

Main Results:

• BUSN achieved a significantly higher median R-squared score (0.9089) compared to the state-of-the-art unsupervised method (0.7153) in body part regression.
• The BUSN pre-processing pipeline improved the mean Dice score for 3D abdominal multi-organ segmentation from 0.7991 to 0.8145.
• The method was trained on 1,030 whole body CT scans and validated on an external cohort of 100 scans.

Conclusions:

• BUSN offers an effective annotation-free approach for body part regression in CT imaging.
• The proposed method enhances spatial localization accuracy and consistency.
• BUSN integration as a pre-processing step improves performance in downstream tasks like 3D organ segmentation.