Master Thesis: Machine Unlearning for Digital Pathology

11.07.2025, Studentische Hilfskräfte, Praktikantenstellen, Studienarbeiten

Machine unlearning is the process of eliminating the effect of a given set of samples, known as the forget set, from a pretrained model. The aim of machine unlearning is to fulfill the "right to be forgotten”, specified in Article 17 of General Data Protection Regulation (GDPR), which grants the individuals such as patients the right to ask data holders including hospitals and medical centers to have their data "forgotten”. "Data forgetfulness” in this regard implies that the data holders are obligated to not only remove the data of the given individuals from their storage systems but also unlearn the contribution of that data from any trained model. The most naive approach to machine unlearning is "exact unlearning”, where the model is retrained from scratch (with randomly initialized weights) on the dataset excluding the forget set. However, this approach is computationally expensive, especially for large (foundation) models and/or frequent unlearning requests. "Approximate unlearning” addresses this challenge by performing the unlearning process in a more computationally efficient fashion.

Previous studies mainly focus on proposing new approximate unlearning algorithms, evaluating the algorithms on non-medical datasets. To the best of our knowledge, the application of machine unlearning on medical data, especially on histopathology images is still underexplored. This project aims to investigate the performance of the existing unlearning algorithms on pathology images, focusing on the patch-level and slide-level classification tasks. Unlike the datasets used in many previous studies, where the samples are considered to be uncorrelated (i.e. belonging to separate individuals/entities), the samples, e.g. patches, in pathology images might belong to the same patient. This characteristic of pathology data makes the unlearning process more challenging. This is because even if a few patients ask for data forgetfulness, the contribution of many patches corresponding to their whole slide images must be unlearned from the model, which might lead to "catastrophic unlearning”, in which the unlearned model might not be generalizable anymore.

Please find the attached PDF for a more detailed description of the project!

Kontakt: reza.nasirigerdeh@tum.de