Generative multi-adversarial network for striking the right balance in abdominal image segmentation

Area of Science:

• Medical Imaging
• Deep Learning
• Computer Vision

Background:

• Semantic segmentation of medical images faces challenges due to class imbalance, where rare abnormalities are difficult to identify.
• Deep learning models struggle to generalize when training data has disproportionately few samples of minority classes.

Purpose of the Study:

• To develop a novel generative multi-adversarial network, Ensemble-GAN, to address class imbalance in abdominal medical image semantic segmentation.
• To improve the generalization and accuracy of identifying rare abnormalities in medical images.

Main Methods:

• Developed Ensemble-GAN, a generative multi-adversarial network with a single generator and multiple discriminators.
• The ensemble model aggregates estimates from multiple models trained on different data subsets and losses.
• Trained and evaluated the framework on the Chaos 2019 and LiTS 2017 public datasets.

Main Results:

• Achieved high F1 scores for segmenting healthy abdominal organs: spleen (0.93), liver (0.96), left kidney (0.90), and right kidney (0.94).
• Demonstrated strong performance in simultaneous segmentation of lesions (F1 score 0.83) and liver (F1 score 0.94).

Conclusions:

• The proposed Ensemble-GAN framework shows superior performance in medical image semantic segmentation compared to existing methods.
• Ensemble-GAN holds significant potential for more accurate abdominal image segmentation, potentially surpassing human expert capabilities.