Semi-Supervised Multi-Organ Segmentation through Quality Assurance Supervision.

Area of Science:

• Medical Image Analysis
• Machine Learning
• Deep Neural Networks

Background:

• Human in-the-loop quality assurance (QA) is crucial for validating medical image segmentation systems.
• Current QA scores are underutilized in machine learning, particularly for training deep neural networks.
• Reusing QA data could improve the efficiency and effectiveness of medical image analysis.

Purpose of the Study:

• To investigate the feasibility of using QA labels as supplementary supervision for semi-supervised learning in medical image segmentation.
• To develop and evaluate a semi-supervised deep neural network incorporating QA scores.
• To enhance the performance of multi-organ segmentation on unlabeled datasets.

Main Methods:

• Proposed a semi-supervised multi-organ segmentation network with a generator and a QA-involved discriminator (pre-trained ResNet-18).
• Trained the generator on 2027 volumes; trained the discriminator on 2D montage images with QA scores (0=success, 1=errors, 2=gross failure).
• Integrated the trained discriminator as a QA-loss function within the segmentation pipeline.

Main Results:

• The discriminator achieved 94% accuracy in classifying QA scores on a test set.
• Incorporating the QA-loss function improved segmentation performance on an unlabeled test dataset, increasing patient data coverage from 714 to 951.
• The number of segmentation failures decreased from 29.90% to 19.83%.

Conclusions:

• QA scores can be effectively utilized as a loss function for semi-supervised learning on unlabeled medical image data.
• A well-trained discriminator can learn from nuanced QA scores, going beyond simple true/false classifications.
• The proposed QA-inspired loss function offers a promising approach for fine-tuning multi-organ segmentation, yielding more robust and accurate results than baseline methods.