BucketAugment: Reinforced Domain Generalisation in Abdominal CT Segmentation.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Computer Vision

Background:

• Deep neural networks excel in medical segmentation but struggle with data outside their training distribution.
• Domain generalization remains a challenge for applying these networks to diverse clinical datasets.

Purpose of the Study:

• Introduce BucketAugment, a novel method to improve domain generalization in CT segmentation using deep neural networks.
• Address the limitations of current deep learning models in handling variations across different medical datasets.

Main Methods:

• BucketAugment utilizes Q-learning principles and validation loss to optimize a policy for 3D volumetric augmentations ('buckets').
• The method employs tunable parameters within these augmentation buckets for flexible integration into existing neural network architectures.

Main Results:

• BucketAugment significantly enhances domain generalization for kidney and liver segmentation across three distinct CT datasets.
• The method demonstrated improved performance with minimal modifications to existing network architectures.

Conclusions:

• BucketAugment offers a promising solution for domain generalization challenges in CT segmentation.
• The approach leverages Q-learning and 3D augmentations to improve deep neural network performance on diverse medical segmentation tasks.