Machine Unlearning-based Privacy-First Medical Imaging Framework for TB Detection
Abstract
Proposes a privacy-first machine unlearning approach for medical imaging, allowing TB detection models to forget patient data on demand while preserving diagnostic performance.
Problem
Deep learning models trained on medical imaging datasets (such as chest X-rays for Tuberculosis detection) achieve high diagnostic accuracy. However, patient privacy regulations require models to support patient data unlearning (the right to be forgotten). Retraining complex convolutional networks or vision transformers from scratch when patients withdraw consent is computationally prohibitive for healthcare networks. The challenge is to quickly erase specific imaging datasets from a trained model without degrading its diagnostic accuracy for remaining patients.
Key Contributions
- A localized machine unlearning algorithm designed specifically for deep convolutional networks.
- A verification protocol evaluating feature residuals after data deletion.
- Benchmarks on chest X-ray databases showing negligible diagnostic degradation.
Methodology
The framework implements a weight-scrubbing update based on Newton step approximations. Instead of full network retraining, we calculate the parameter updates by computing the Hessian matrix of the loss function over a small subset of remaining data. This is optimized using lower-bound projections to keep memory overhead low. To verify that the scrubbed patient images are fully purged, we compute representation similarity metrics in the latent space of the network, confirming that the network activations for unlearned samples are indistinguishable from random noise.
Results
Unlearning chest X-ray samples from a ResNet-based TB detection model was completed in under 4.2 seconds—representing a massive improvement over the hours required for full retraining. The diagnostic accuracy (AUC-ROC) for TB classification remains stable at 94.6% before and after the unlearning operation.
Implications
Verifiable and efficient unlearning algorithms enable privacy-first diagnostic systems in hospital networks. Patient consent can be managed dynamically, allowing institutions to strictly comply with privacy mandates without losing utility in their trained AI systems.