Longitudinal Lesion Inpainting in Brain MRI via 3D Region Aware Diffusion

A research team from institutions including Concordia University and McGill University has published a breakthrough AI model for medical imaging. Their novel framework, Longitudinal Lesion Inpainting via 3D Region Aware Diffusion, tackles a critical problem in neurology: evolving lesions in brain MRI scans (from conditions like multiple sclerosis) can bias automated analysis tools, obscuring the true progression of neurodegenerative disease. The team's solution is a pseudo-3D Denoising Diffusion Probabilistic Model (DDPM) that uses multi-channel conditioning. It intelligently incorporates context from a patient's previous scans (longitudinal data) to inpaint, or fill in, only the pathological lesion areas in a new scan, leaving surrounding healthy brain tissue completely untouched.

This "Region-Aware Diffusion" (RAD) mechanism is key to its performance and efficiency. Evaluated on longitudinal MRI data from 93 patients, the model significantly outperformed the leading previous method, FastSurfer-LIT. It achieved a 57% improvement in perceptual fidelity, lowering the Learned Perceptual Image Patch Similarity (LPIPS) score from 0.07 to 0.03, and produced superior 3D anatomical continuity without inter-slice discontinuities. Most impressively, the RAD focus allows for massive computational gains. The framework processes an entire 3D brain volume in just 2.53 minutes on average, representing a roughly 10x speedup over LIT's 24.30 minutes. It also demonstrates high longitudinal stability with a Temporal Fidelity Index of 1.024, close to the ideal 1.0.

The model's primary impact is as a highly reliable and efficient preprocessing step. By generating "lesion-free" versions of MRI scans, it removes a major source of noise for downstream automated pipelines that measure brain volume, cortical thickness, or other biomarkers. This allows clinicians and researchers to track the subtle, true progression of diseases like Alzheimer's or MS with greater accuracy, potentially leading to better patient monitoring and therapeutic assessment. The team plans to release their code and a derivative dataset of 93 pre-processed scans upon the paper's acceptance.