Outlier Guided Optimization of Abdominal Segmentation.

Area of Science:

• Medical Image Analysis
• Artificial Intelligence in Healthcare
• Radiology

Background:

• Abdominal multi-organ segmentation in computed tomography (CT) is challenging due to anatomical variability.
• Scaling datasets for improved segmentation is costly and often impractical.
• The marginal value of additional data for segmentation models is not well understood.

Purpose of the Study:

• To propose a single-pass active learning method using human quality assurance (QA) for abdominal multi-organ segmentation.
• To evaluate the effectiveness of augmenting datasets with outlier versus inlier data.
• To determine the optimal strategy for data selection to enhance segmentation performance.

Main Methods:

• Utilized a pre-trained 3D U-Net model for abdominal multi-organ segmentation.
• Augmented the dataset with either outlier (failed cases) or inlier (successful cases) data.
• Trained new models using augmented datasets with 5-fold cross-validation and withheld samples.

Main Results:

• Manual labeling of outliers increased Dice scores by 0.130, compared to 0.067 for inliers (p<0.001).
• Adding outliers provided higher marginal value than adding inliers for improving segmentation.
• Performance gains were achieved without compromising multi-organ segmentation or significantly increasing training time.

Conclusions:

• Identifying and correcting baseline algorithm failures is an effective and efficient method for selecting training data.
• Active learning focused on outliers enhances abdominal multi-organ segmentation performance.
• This approach offers a practical solution for improving medical image segmentation models.