Low Data Image Analysis with a Generative Adversarial Network: A Case Study on Women Pelvic MRI Scans

Medical image analysis and geometry reconstruction have vast applications in evaluating pathology and treatment. The emergence of machine learning (ML) has led to considerable progress in the field. However, the performance of ML algorithms strongly depends on the size of the training dataset. Providing quality data, though, is often expensive, laborious, and time-consuming. The approach presented in this paper demonstrates that Generative Adversarial Networks (GANs) are helpful for generating high-quality synthetic images. With the artificially generated data, the size of the required clinical images for 3D reconstruction of patient-specific pelvic can remain small. Our results show that given just five clinical MRI scans, the ML performs image segmentation with %68 accuracy. This number increases to %85 when synthetic data generated by the GAN model is added to the training set. Details on the code for the generative model are publicly available at https://github.com/sabino-ramirez/task driven data augmentation.git