Ensembling Low Precision Models for Binary Biomedical Image Segmentation.

Area of Science:

• Medical image analysis
• Computer-aided diagnosis
• Biomedical imaging

Background:

• Accurate segmentation of anatomical regions and lesions in medical images is challenging due to similar foreground/background appearances.
• Automatic segmentation algorithms often produce asymmetric errors, with more false positives than false negatives.

Purpose of the Study:

• To leverage the asymmetry in segmentation errors to improve automatic segmentation performance.
• To develop a generalizable strategy applicable to various medical imaging modalities and segmentation tasks.

Main Methods:

• Training a diverse ensemble of segmentation models with intentionally high recall and low precision.
• Aggregating predictions from the ensemble to cancel out individual models' false positive errors.
• Applying the strategy to carotid artery, myocardium, and multiple sclerosis lesion segmentation.

Main Results:

• The proposed ensemble strategy significantly boosts performance over baseline segmentation methods.
• True positive segmentations are consistent across models, while false positive errors are diverse and tend to cancel out.
• Demonstrated effectiveness across CT angiography, cardiovascular MRI, and brain MRI.

Conclusions:

• Leveraging asymmetric errors by combining high-recall, low-precision models is an effective strategy for improving medical image segmentation.
• The method offers a generalizable approach to enhance segmentation accuracy in diverse clinical applications.
• This technique addresses a key challenge in automated medical image analysis, reducing reliance on manual segmentation.