Lesion-Enhanced Fast MRI with Weakly Supervised, Model-Driven Deep Learning

Motivation: Most research on accelerated MRI techniques focuses on improving overall image quality at the voxel level, neglecting specific abnormalities that are clinically significant. Goal(s): This work achieves accelerated MRI reconstruction in a weakly supervised setting while also localizing the lesion areas and improving the reconstruction quality at those lesion locations. Approach: We built a task-specific MRI reconstruction model that includes customized learnable regularization, which is solved by unfolding into a network using alternating optimization. Results: Tests on public medical datasets show that our method significantly outperforms current benchmark approaches and demonstrates substantial improvements in the field of pathology. Impact: In a weakly supervised setting, our method uses only MRI image-level labels to achieve accelerated MRI reconstruction while localizing the lesion areas and improving their reconstruction quality. This has significant implications for clinical applications.