Generative AI for Artifact Correction and Privacy-Secure Medical Imaging
Time: Fri 2024-12-13 13.00
Location: F2, Lindstedtsvägen 26 & 28, Stockholm
Video link: https://kth-se.zoom.us/j/69355780837
Language: English
Subject area: Computer Science
Doctoral student: Lennart Alexander Van der Goten , Beräkningsvetenskap och beräkningsteknik (CST), Science for Life Laboratory, SciLifeLab
Opponent: Associate Professor Ida Häggström, Chalmers University of Technology, Gothenburg, Sweden
Supervisor: Associate professor Kevin Smith, Science for Life Laboratory, SciLifeLab, Beräkningsvetenskap och beräkningsteknik (CST); Professor Zeynep Akata, Technical University of Munich, Munich, Bavaria, Germany
QC 20241120
Abstract
Magnetic Resonance Imaging (MRI) is a widely used non-invasive technology that provides detailed visualizations of internal body structures, particularly soft tissues such as the brain, muscles, and internal organs. However, despite its crucial role in modern healthcare, MRI faces significant obstacles that limit its effectiveness. One major challenge is that MRI scans can unintentionally reveal identifiable facial features, creating potential privacy risks if this information is misused for re-identification. This raises serious concerns about data security in an increasingly digital healthcare landscape. Additionally, MRI scans require patients to remain still during imaging, as even slight movements can degrade image quality or, in severe cases, render scans unusable, leading to costly re-scans and patient discomfort.
To address these challenges, this thesis leverages generative modeling techniques using artificial neural networks. For the first challenge, we introduce a novel data-driven remodeling-based approach to visually de-identify MRI scans while preserving medically relevant information, such as the brain. Conventional methods that remove sensitive regions (e.g. the face or ears) often disrupt downstream analysis by introducing a domain shift—a significant alteration in data distribution that hampers diagnostic accuracy. Our approach generates a realistic remodeling of these sensitive areas, maintaining privacy while preserving diagnostic utility and downstream task performance.
For the second challenge, we develop techniques to remove artifacts from MRI scans, allowing the recovery of scans that would otherwise be unusable. By integrating 3D vision transformers with self-supervised and transfer learning, our methods enhance image quality while minimizing computational cost. This reduces the need for re-scanning, improves diagnostic accuracy, and enhances patient comfort by streamlining the MRI process.
Our findings highlight the transformative potential of generative modeling in medical imaging. By addressing both privacy risks and artifact removal, this research establishes new standards for secure, efficient, and precise diagnostics. With the growing integration of AI in healthcare, these innovations lay the groundwork for scalable, privacy-conscious, and accessible diagnostic practices across various imaging modalities.