Multi-structure Segmentation from Partially Labeled Datasets. Application to Body Composition Measurements on CT...

Area of Science:

• Medical Image Analysis
• Deep Learning
• Computer-Aided Diagnosis

Background:

• Labeled data is a significant bottleneck in medical image research, particularly for multi-organ segmentation.
• Existing multi-organ segmentation frameworks struggle with datasets containing multiple label maps of individual structures from different cases.
• The need for efficient methods to leverage available, often partially annotated, medical imaging data is critical.

Purpose of the Study:

• To develop a robust multi-organ deep learning segmentation network using partially labeled medical images.
• To propose a modified cost function and network architecture that effectively utilizes segmentations from multiple organs across different cases.
• To evaluate the proposed methods against traditional approaches using chest CT scans.

Main Methods:

• A modified cost function was developed to focus on labeled voxels, ignoring unlabeled structures.
• A Convolutional U-Net (CUNet) architecture was proposed, featuring added convolutions in skip connections.
• Performance was evaluated on pectoralis muscle and subcutaneous fat segmentation in chest CT scans, comparing PUNet, multiple UNets, multi-class UNet, and CUNet.

Main Results:

• Training with partially labeled images (PUNet) showed performance equivalent to a multi-class UNet trained on fully labeled data (Dice coefficients of 0.909 vs 0.906).
• The proposed CUNet architecture, trained with partially labeled images, significantly outperformed other methods with a Dice coefficient of 0.916 (p<0.0001).
• The study demonstrated that partial labeling is as effective as full labeling when annotation counts are constant, and CUNet enhances performance.

Conclusions:

• Partially labeled data can be effectively utilized for training multi-organ segmentation networks, achieving performance comparable to fully labeled datasets.
• The proposed CUNet architecture, incorporating convolutions in skip connections, offers a significant performance improvement for medical image segmentation.
• This research provides a viable solution to the labeled data bottleneck in medical image analysis, enabling more robust and efficient deep learning models.