BEGIN:VCALENDAR
VERSION:2.0
PRODID:-//132.216.98.100//NONSGML kigkonsult.se iCalcreator 2.20.4//
BEGIN:VEVENT
UID:20260523T215619EDT-8464GaisRw@132.216.98.100
DTSTAMP:20260524T015619Z
DESCRIPTION:Abstract\n\nEnsuring the trustworthiness of deep learning in me
 dical imaging requires going beyond accuracy to address explainability and
  fairness. In safety-critical domains such as healthcare\, clinical adopti
 on depends not only on predictive performance but also on a model’s abilit
 y to provide transparent reasoning. This thesis introduces generative mode
 ling approaches\, including Vision-Language Foundation Models (VLFMs)\, to
  synthesize counterfactual (CF) images\, hypothetical alternatives that si
 mulate the effect of modifying specific clinical attributes thereby enabli
 ng both interpretability and bias mitigation.\n\nWe first address the gap 
 that predictive models are not designed to discover personalized imaging m
 arkers predictive of future outcomes. To this end\, we develop the first d
 eep conditional 3D generative model\, based on GANs\, for subject-specific
  counterfactual generation in brain MRI of patients with relapsing-remitti
 ng multiple sclerosis. By analyzing differences between factual and CF ima
 ges\, local modifications predictive of future disease activity can be ide
 ntified\, enabling interpretable biomarker discovery. Validation on a larg
 e\, multi-center dataset shows alignment with established markers of progr
 ession. While this model provided the first subject-specific counterfactua
 ls for volumetric imaging\, it was still limited latching onto the feature
 s of the majority class\, and thereby biased. This led to the second contr
 ibution focused on fairness.\n\nNext\, we address the issue that medical i
 maging classifiers often latch onto spurious correlations instead of disea
 se markers\, yielding biased models that are not ‘right for the right reas
 ons.’ To overcome this\, we propose the first end-to-end framework that in
 tegrates debiasing with counterfactual explanations using Cycle-GAN. By co
 mbining Empirical Risk Minimization (ERM) and Group-DRO with counterfactua
 l generation\, the framework disentangles disease features from shortcut s
 ignals and introduces the Spurious Correlation Latching Score (SCLS) to qu
 antify model reliance on such biases. Experiments on CheXpert and RSNA dat
 asets with both synthetic and real artifacts demonstrate improved generali
 zation across underrepresented subgroups. Nonetheless\, mitigation remaine
 d constrained by the limited generative capacity of Cycle-GAN\, motivating
  the development of DeCoDEx—a diffusion-based counterfactual generator gui
 ded by an artifact detector that enables manipulation of disease markers w
 hile preserving clinically relevant features. However\, this method still 
 lacks the precision and resolution required for fine-grained clinical edit
 ing.\n\nTo overcome this limitation\, we introduce PRISM\, the first frame
 work to adapt vision–language foundation models (VLFMs) to medical imaging
  for high-resolution\, language-guided counterfactual synthesis. PRISM ena
 bles fine-grained editing of 2D medical images\, such as the selective rem
 oval or insertion of devices (e.g.\, pacemakers\, wires)\, while faithfull
 y preserving anatomical structures. By leveraging natural language prompts
 \, PRISM generates clinically meaningful counterfactuals that improve the 
 fairness and robustness of downstream classifiers and provide data augment
 ation for underrepresented groups. Beyond targeted edits\, PRISM also lays
  the groundwork for future research into uncovering hidden attribute–image
  relationships: we demonstrate that VLFM models can be queried with textua
 l prompts to reveal associations between visual features and clinical attr
 ibutes. Finally\, we extend this research work by showing how reinforcemen
 t learning can be used to better align text-to-image generation with clini
 cal relevance. We also demonstrate that VLFMs can disentangle key factors 
 of variation\, such as disease severity or the presence of artifacts\, the
 reby enabling more controllable and interpretable medical image synthesis.
 \n
DTSTART:20251030T153000Z
DTEND:20251030T173000Z
LOCATION:Room 603\, McConnell Engineering Building\, CA\, QC\, Montreal\, H
 3A 0E9\, 3480 rue University
SUMMARY:PhD defence of Amar Kumar – Explainability and Fairness in Medical 
 Imaging via Counterfactual Generation
URL:/ece/channels/event/phd-defence-amar-kumar-explain
 ability-and-fairness-medical-imaging-counterfactual-generation-368189
END:VEVENT
END:VCALENDAR