Potential Pitfalls of Deep Learning in Medical Image Reconstruction
Abstract:
This session will focus on scenarios in which deep learning algorithms developed for medical imaging might produce unreliable results, e.g. due to distribution shifts, bias, hallucinations, or other factors. The session is planned following the increasing interest in studying sensitivities and instabilities of such algorithms. The talks will discuss strategies for exposing algorithmic sensitivities and addressing them, preventing inverse crimes, and increasing algorithmic interpretability. The aim of the session is to raise awareness to the growing problem of unreliable AI performance in the context of medical imaging, suggest guidelines and solutions, and invoke community discussions.
Organisers: Efrat Shimron (Chair) & Florian Knoll (Committee member)
Session Schedule
- 17:30 - 17:50 Invited talk -
- The Risk in Naive Training of Medical AI Algorithms: Pitfalls, Stability and Generalizability Issues
- Efrat Shimron
- 17:50 - 18:10 Invited talk -
- Trustworthy Image Reconstruction and beyond: Using Adversarial Approaches
- Daniel Rueckert
- 18:10 - 18:30 Invited talk -
- Generalizability (or not?) of patch-based image models
- Jeffrey Fessler
- 18:30 - 18:50 Invited talk -
- The importance of signal modeling for MRI artifact mitigation
- Jon Tamir
- 18:50 - 19:10 Invited talk -
- Generalised hardness of approximation and hallucinations -- On barriers and paradoxes in image reconstruction
- Anders Hansen
- 19:10 - 20:30 Invited poster -
- Mitigating Data Paucity and Distributional Shifts for Accelerated MRI Alongside New Clinically-Relevant Evaluation Metrics
- Akshay S. Chaudhari
- 19:10 - 20:30 Contributed poster -
- Physics-Driven Data Priors for Robust Self-Supervised Accelerated MRI Reconstruction
- Arjun Desai*
- 19:10 - 20:30 Contributed poster -
- MRI Reconstruction via Data-Driven Markov Chains with Joint Uncertainty Estimation
- Guanxiong Luo*
- 19:10 - 20:30 Contributed poster -
- Partial privacy loss in machine learning: a statistical signal processing perspective
- Tamara T. Mueller*
- 19:10 - 20:30 Committee member poster -
- Insights into the reliability of deep learning reconstructions with research challenges
- Florian Knoll