IBIIS-AIMI Seminar - Cynthia Xinran Li, PhD Student & Magdalini Paschali, PhD

Speakers: 

Cynthia Xinran Li, PhD Student: Institute for Computational and Mathematical Engineering, Stanford University 

Talk Title: Improving Prostate Cancer Detection on MRI using Anatomical Prior Knowledge and Multi-Teacher Distillation

Abstract: MRI is increasingly being used to detect prostate cancer, yet its interpretation can be challenging due to subtle differences between benign and cancerous regions. Many deep learning models have been developed to automatically detect prostate cancer, yet these methods can be further improved by incorporating useful anatomical information about prostate tumors (e.g. ~70% of prostate cancers occur in the peripheral zone). We quantified the heterogeneity of prostate cancer occurrence using a registration pipeline to compute voxel-level cancer distribution and incorporated it as a prior distribution for a Denoising Diffusion Probabilistic Model (DDPM). Our method achieved statistically significantly higher accuracy and Dice coefficient than alternative models. As many AI methods have been developed for similar problems, a selective feature ensemble method is needed to combine their strength and eliminate incorrect predictions. We developed a multi-teacher distillation pipeline with a reinforcement-learning agent to dynamically determine distillation weights, thereby enriching the student model with useful knowledge and outperforming all teacher models.

 

Magdalini Paschali, PhD: Postdoctoral Scholar, Department of Radiology, Stanford University 

Talk Title: Controllable 3D Medical Image Synthesis for Validating and Stress-Testing Clinical AI

Abstract: The rigorous evaluation and stress-testing of clinical AI models is frequently constrained by a lack of diverse, precisely controlled data, especially for rare pathologies or subtle clinical edge cases. This talk presents a framework for generating high-fidelity, controllable 3D CT volumes by conditioning latent diffusion models on clinical text, anatomical segmentations, and demographics to create targeted evaluation datasets. We demonstrate this approach across two distinct domains: generating 3D chest CTs conditioned on radiology reports to simulate specific findings like pulmonary embolism and medical devices and creating counterfactual head CTs to model varied manifestations of intracranial hemorrhage. By evaluating these synthetic volumes through downstream diagnostic tasks and sensitivity analyses, we show that synthetic data can effectively preserve clinically relevant signals and provide a path for the rigorous evaluation and stress-testing of AI models before clinical deployment.
 

Attendance is open to the Stanford and AIMI affiliate community. Please contact aimicenter@stanford.edu for the Zoom link if you would like to attend virtually.