Multiorgan segmentation from partially labeled datasets with conditional nnU-Net.

Area of Science:

• Medical Imaging and Computer Vision
• Deep Learning for Medical Image Analysis

Background:

• Multiorgan abdominal CT segmentation is crucial for clinical applications like treatment planning.
• Existing methods require fully annotated data, which is labor-intensive and time-consuming to acquire.
• Partially labeled datasets are more readily available but underutilized by current segmentation networks.

Purpose of the Study:

• To develop a novel deep learning model for multiorgan abdominal CT segmentation using partially labeled datasets.
• To overcome the limitations of fully supervised methods by leveraging readily accessible, albeit incomplete, annotations.
• To improve the efficiency and robustness of organ segmentation in CT imaging.

Main Methods:

• Proposed a conditional nnU-Net architecture, utilizing the state-of-the-art nnU-Net as a backbone.
• Introduced a conditioning strategy to incorporate auxiliary information into the decoder, aiding pixel-wise organ identification and spatial information recovery.
• Implemented deep supervision for multi-scale output refinement and combined Dice and Focal loss for optimized training.

Main Results:

• Achieved promising segmentation performance across seven publicly available CT datasets (liver, pancreas, spleen, kidney).
• Demonstrated the effectiveness of the conditional approach in leveraging partially labeled data for improved segmentation.
• Successfully integrated non-overlapping labeled datasets, mitigating the issue of data scarcity in multiorgan segmentation.

Conclusions:

• The conditional nnU-Net effectively utilizes partially labeled datasets for robust multiorgan abdominal CT segmentation.
• The proposed method alleviates the data hunger problem and breaks down barriers between disparate labeled datasets.
• This approach offers a promising solution for improving the accessibility and efficiency of medical image segmentation.