Corrective Unlearning for MRI Reconstruction

Sairam Narendra Babu, Aryaman Bahl, Chinmay Sharma, Ponnurangam Kumaraguru, Balaji Darur

Keywords: Medical Imaging, Corrective Machine Unlearning, MRI Reconstruction, Deep Learning, Selective Synaptic Dampening, Trustworthy AI, MRI Acceleration

TL;DR: Applying Corrective Unlearning techniques to bolster the robustness of Deep Learning based MRI Reconstruction models to possibly poisoned data

Abstract:

Magnetic Resonance Imaging reconstruction accelerates image acquisition by reconstructing high-quality images from undersampled k-space data using deep learning. However, real-world deployment of these models remains hindered by concerns around trustworthiness, generalization, and data privacy, especially in the presence of corrupted or adversarial training samples. We propose a Corrective Machine Unlearning framework that selectively removes the influence of harmful data while preserving overall model performance. By leveraging techniques such as Selective Synaptic Dampening, our approach aims to robustly and efficiently forget poisoned representations. Experimental results on MRI reconstruction tasks demonstrate that Corrective Machine Unlearning can effectively mitigate artifacts introduced through data poisoning while maintaining high fidelity on untainted inputs. Our findings underscore the promise of corrective unlearning as a practical step toward safer, privacy preserving, and clinically reliable MRI systems. All code and scripts used are available at [Github repository](https://anonymous.4open.science/r/CorrectiveMachineUnlearningForMRI-0C22/)
.

Submission Number: 18