Data augmentation using generative adversarial networks (CycleGAN) to improve generalizability in CT segmentation...

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Computer Vision

Background:

• Labeled medical imaging data is scarce and costly, hindering the development of generalizable deep learning models.
• Data augmentation is crucial for improving model generalizability, with generative adversarial networks (GANs) offering a novel approach.

Purpose of the Study:

• To evaluate the effectiveness of CycleGAN for data augmentation in computed tomography (CT) segmentation tasks.
• To assess the impact of CycleGAN-generated synthetic non-contrast CT images on segmentation performance, particularly on out-of-distribution data.

Main Methods:

• Trained a CycleGAN to translate contrast-enhanced CT images into non-contrast CT images.
• Augmented a training dataset with synthetic non-contrast images generated by CycleGAN.
• Compared the segmentation performance of U-Net models trained on original versus augmented datasets using in-distribution (contrast CT) and out-of-distribution (non-contrast CT) datasets.

Main Results:

• CycleGAN augmentation significantly improved segmentation performance on out-of-distribution non-contrast CT data, notably for kidney segmentation (Dice score increased from 0.09 to 0.66).
• Improvements were also observed for liver (0.86 to 0.89) and spleen (0.65 to 0.69) segmentation.
• The method demonstrated enhanced generalizability of deep learning models to unseen CT imaging data.

Conclusions:

• CycleGAN is a valuable tool for data augmentation in CT segmentation, effectively improving model performance on out-of-distribution datasets.
• This approach can reduce the need for extensive manual segmentation, lowering costs and effort in medical imaging research.
• The findings suggest broader applicability of GAN-based augmentation for enhancing deep learning in medical imaging.