Marginal loss and exclusion loss for partially supervised multi-organ segmentation

Area of Science:

• Medical Imaging
• Computer Vision
• Machine Learning

Background:

• Multi-organ segmentation in medical images is crucial but hindered by costly and time-consuming annotation processes.
• Existing datasets are often small and partially labeled, presenting challenges for training robust segmentation networks.

Purpose of the Study:

• To develop a method for training a single multi-organ segmentation network using a union of partially labeled datasets.
• To introduce novel loss functions specifically designed for learning from incomplete multi-organ annotations.

Main Methods:

• Proposed two novel loss functions: marginal loss and exclusion loss.
• Marginal loss incorporates the probability of a 'merged' background label, suitable for partially labeled data.
• Exclusion loss measures dissimilarity between labeled and unlabeled organs, leveraging their non-overlapping nature.

Main Results:

• Demonstrated significant performance improvements on a union of five benchmark multi-organ segmentation datasets.
• The proposed loss functions enhanced state-of-the-art methods without increasing computational overhead.
• Effective for segmenting organs including the liver, spleen, kidneys, and pancreas.

Conclusions:

• Novel marginal and exclusion loss functions effectively address the challenge of training multi-organ segmentation networks on partially labeled data.
• The proposed approach offers a computationally efficient way to improve segmentation accuracy by leveraging diverse, incomplete datasets.
• This work provides a valuable contribution to medical image analysis, enabling more robust and accurate organ segmentation.