Generative Augmentation for Brain Adenoma MRI Segmentation

Richard Szarka

Supervisor(s): prof. Ing. Vanda Benešová, Phd.

Slovak Technical University

Abstract: Deep learning has shown great promise in fields such as radiology, but training robust segmentation networks is difficult in small healthcare systems due to the lim- ited availability of annotated MRI datasets. To overcome this challenge, we propose a generative data augmentation approach that enhances training efficiency by synthesiz- ing both MRI scans and their corresponding segmentation masks. Our method consists of two key components: the Key- Slice Extractor Network and the Conditional Diffusion Projected Generative Adversarial Network (CDP-GAN). The Key-Slice Extractor Network is designed to identify the most relevant MRI slices, reducing redundancy and improving the quality of the training dataset. These se- lected slices serve as inputs for our CDP-GAN, which utilizes slice index conditioning and diffusion-based aug- mentation to generate realistic 3D MRI volumes of pi- tuitary adenoma (PA). Our model produces anatomically consistent synthetic MRI scans and generates correspond- ing segmentation masks, enabling a direct expansion of labelled training data. To assess the impact of our approach, we evaluate the performance of a 3D U-Net segmentation model trained on different combinations of real and synthetic MRI data. Our findings suggest that incorporating CDP-GAN-generated samples, complete with segmentation masks, should en- hance segmentation accuracy, demonstrating the effective- ness of generative augmentation in data-scarce settings. This framework offers a scalable solution for improving AI-driven medical imaging applications, particularly in re- gions with limited access to large annotated datasets.
Keywords: Computer Vision
Full text:
Year: 2025