Generative AI for Multi-Channel Pathological Brain MRI Synthesis

NAISS 2024/5-51


NAISS Medium Compute

Principal Investigator:

Robin Strand


Uppsala universitet

Start Date:


End Date:


Primary Classification:

20603: Medical Image Processing

Secondary Classification:

10207: Computer Vision and Robotics (Autonomous Systems)




Medical imaging, particularly Magnetic Resonance Imaging (MRI), is pivotal in diagnosing neurological disorders. However, challenges arise due to the scarcity and privacy concerns surrounding annotated pathological brain MRI datasets, impacting research and machine learning model training. This project proposes a solution by applying Generative Artificial Intelligence (AI) to synthesize diverse multichannel brain MRI images with various pathologies. Diffusion models, particularly those based on neural networks, have proven effective in capturing intricate patterns in medical images. Our project integrates these models into the synthesis of realistic multichannel brain MRI images. This technique allows for the exploration of latent spaces, facilitating the generation of images that not only resemble real-world pathologies but also exhibit controlled variations crucial for training models to handle clinical heterogeneity. The generative AI framework will be trained on existing MRI datasets containing labeled pathological images. Through this training process, the model learns the complex relationships between normal and abnormal brain structures, enabling the generation of synthetic multichannel images that accurately reflect the nuances of different pathologies. The synthesized multichannel dataset produced by our approach serves as a valuable resource for training machine learning models, alleviating the challenges associated with obtaining annotated data for various pathologies. Additionally, it contributes to augmenting existing datasets, enhancing the robustness of AI algorithms in handling diverse clinical scenarios. This project addresses the challenges of data scarcity and privacy concerns, ultimately improving the accuracy and generalization capabilities of diagnostic models in neurology. The project's outcomes will undergo rigorous quantitative and qualitative assessments. We will compare the performance of models trained on synthesized multichannel data with those trained on real-world datasets. This evaluation aims to provide insights into the efficacy of our generative AI approach in replicating complex pathological patterns and supporting the development of reliable diagnostic tools for healthcare practitioners. In conclusion, our project leverages the synergy between generative AI and diffusion models to synthesize realistic multichannel brain MRI images with pathologies. By addressing challenges related to limited and privacy-sensitive datasets, the proposed approach has the potential to advance the field of medical imaging, improving the accuracy and generalization capabilities of diagnostic models in neurology, benefiting both researchers and healthcare practitioners.