Zhang, Jinwei; Zuo, Lianrui; Liu, Yihao; Zhang, Hang; Remedios, Samuel W.; Landman, Bennett A.; Calabresi, Peter A.; Saidha, Shiv; Newsome, Scott D.; Pham, Dzung L.; Prince, Jerry L.; Mowry, Ellen M.; Carass, Aaron. (2027).听.听Medical Image Analysis, 115, 104338.听
Multiple sclerosis (MS) lesions can interfere with automated MRI analysis, while limited labeled data make it difficult to train models that accurately detect these lesions. We developed MSRepaint, an AI-based imaging method that can both fill lesion-affected areas with realistic healthy-appearing tissue and generate synthetic lesions for training purposes. The method preserves surrounding brain anatomy and can work when some MRI inputs are unavailable. In testing, MSRepaint outperformed established lesion-filling methods and achieved accuracy comparable to a recent AI-based approach while processing images more than 20 times faster. Models trained using its synthetic lesions also performed better across multiple lesion-detection datasets than models trained with other synthetic or real training data. MSRepaint therefore provides a fast, flexible approach for improving MRI analysis and generating realistic training data for MS lesion detection.
Fig. 1.听Overview of MSRepaint training and inference. (a) Training data: A 2D multicontrast image聽饾懃0聽undergoes forward diffusion, generating noisy images聽饾懃饾憽鈭�1,聽饾懃饾憽, and the final degraded image聽饾懃饾憞, paired with its lesion mask聽饾憖. (b) Training: The denoiser聽饾湙饾懁聽is trained to estimate聽藛饾懃饾憽鈭�1聽from聽饾懃饾憽聽with guidance from mask聽饾憖. (c) Inference: The trained denoiser聽饾湙饾懁鈭椔爂enerates outputs guided by聽饾憖target, while the repaint mask聽饾憖repaint聽restricts updates to its region, preserving surrounding tissue. A repainting mechanism iteratively refines update boundaries to ensure seamless integration with the surrounding tissue. (d) Example tasks: Lesion filling sets聽饾憖target=0聽to replace lesions with healthy tissue, while lesion synthesis sets聽饾憖target=饾憖repaint聽to generate new lesion patterns.