[CCoE Notice] Dissertation Announcement: Samira Zare, "Invariant and Transformer-Based Learning for Reducing Biases in Medical AI"

[Greenwell, Stephen J]

[Dissertation Defense Announcement at the Cullen College of Engineering]
INVARIANT AND TRANSFORMER-BASED LEARNING FOR
REDUCING BIASES IN MEDICAL AI
Samira Zare
November 18, 2024; 11 a.m. - 1 p.m. (CST)
Zoom: https://urldefense.com/v3/__https://us06web.zoom.us/j/4055667438?pwd=T0tLZHdSSGxQWkJRYVIxZ1ljbmFjQT09__;!!LkSTlj0I!D6Y8xY9biuinWtpP12tvN3XKb3rQVX45Bs7ljyjuyQ5eAcPMkGgGvpaFuTjRVavewYqm87ySj-ki8S5LXDyjQwzUCjE$ 

Committee Chair:
Hien V. Nguyen, Ph.D.
Committee Members:
Saurabh Prasad, Ph.D. | David Mayerich, Ph.D. | Aaron Becker, Ph.D. | Daniel Floryan, Ph.D.
Abstract
Deep learning has revolutionized medical image analysis, demonstrating exceptional capabilities in the interpretation of complex medical images, including X-rays, MRI scans, and CT scans. These advancements have enabled earlier and more accurate diagnoses of diseases like cancer, fractures, and neurological disorders. However, challenges such as selection biases, confounding variables, and poor generalization across diverse datasets limit the reliability of these models in healthcare applications. This dissertation addresses these challenges by developing novel approaches to enhance model robustness in medical image analysis.

We propose a dynamic set-operator using Transformer architecture to aggregate prognostic features from nephropathology images, ensuring permutation invariance and improved generalization. Furthermore, we investigate invariant learning algorithms to mitigate biases in medical image classification, focusing on overcoming the limitations of traditional Empirical Risk Minimization (ERM). Our experiments demonstrate that invariant learning, combined with environment discovery methods, can enhance the robustness of models against spurious correlations and confounding variables.

Additionally, we explore the potential of large language models (LLMs) in assisting primary care physicians with real-time diagnostic support, examining their effectiveness and biases across different demographic groups. The findings underscore the potential benefits along with the limitations for the safe deployment of AI in healthcare.

In conclusion, this research contributes to the development of deep learning models that can generalize effectively across different medical domains while minimizing biases. By leveraging advanced algorithms, invariant learning techniques, and large pre-trained models, we aim to pave the way for more reliable and equitable AI applications in healthcare.
[Engineered For What's Next]


-------------- next part --------------
An HTML attachment was scrubbed...
URL: http://Bug.EGR.UH.EDU/pipermail/engi-dist/attachments/20241112/27e0b63c/attachment.html