Can Generative Adversarial Networks help to overcome the limited data problem in segmentation?

Area of Science:

• Medical image analysis
• Deep learning in radiology

Background:

• Deep learning models like Generative Adversarial Networks (GANs) show promise in image translation.
• U-Net architectures are traditional deep learning models for segmentation tasks.

Purpose of the Study:

• To investigate if GANs offer a performance boost over U-Net for medical image segmentation with small training datasets.
• To compare the efficacy of U-Net and conditional GAN (cGAN) models in segmentation tasks with limited data.

Main Methods:

• Trained U-Net and U-Net with patch discriminator (cGAN) models on varying dataset sizes (1-100 patients).
• Evaluated segmentation performance for male pelvis CT data using Dice similarity coefficient and Hausdorff distance.
• Assessed model performance in relation to training dataset size.

Main Results:

• No significant performance difference was observed between U-Net and cGAN models when trained with identical dataset sizes up to 100 patients.
• Both models demonstrated significant performance improvements as training dataset size increased from 1 to 20 patients.
• The size of the training dataset had a more substantial impact on performance than the choice of model architecture.

Conclusions:

• Conditional GANs did not provide a significant performance advantage over U-Net for the tested segmentation task, even with small datasets.
• Increasing the training dataset size is crucial for improving segmentation model performance.
• The study highlights the importance of data quantity over advanced architectures like GANs when dealing with limited medical imaging data.