Department of Computational and Data Sciences
Department Seminar
Speaker : Mr. Vivek Gopalakrishnan, Ph.D. Student @ MIT
Title : “Computer Vision in the Operating Room: Differentiable X-ray Rendering for Intraoperative Inverse Problems”
Date & Time : January 13, 2025 (Monday), 02:00 PM
Venue : # 102, CDS Seminar Hall
ABSTRACT
Humans possess the innate ability to perceive the 3D geometry of scenes in natural images, inspiring numerous attempts in computer vision to build machines capable of doing the same. However, for medical images, the complexity of human anatomy, combined with optical properties inherent to medical imaging systems, limits the accuracy of our mental reconstructions (e.g., the penetrating radiation used in X-rays makes it impossible to determine which objects occlude others from imaging alone). In this talk, I will demonstrate how differentiable implementations of the X-ray image formation model can power self-supervised solutions to inverse problems in intraoperative settings. Our key contributions include (A) DiffPose, a framework for training patient-specific 2D/3D X-ray to CT registration networks that achieve, for the first time, consistent submillimeter accuracy and (B) DiffVox, a 3D CBCT reconstruction method that achieves state-of-the-art 3D reconstructions from sparse numbers of posed X-ray images. Finally, I will briefly demonstrate algorithmic advancements towards making these computer vision methods surgically deployable.
- DiffDRR (rendering): https://github.com/eigenvivek/DiffDRR
- DiffPose (registration): https://arxiv.org/abs/2312.06358
- DiffVox (reconstruction): https://arxiv.org/abs/2411.19224
BIO: Vivek is a PhD student in Medical Engineering and Medical Physics at the Harvard-MIT Program in Health Sciences and Technology, advised by Polina Golland. The goal of his research is to address unmet clinical needs through the development of biomedical machine learning methods that deepen our ability to understand and treat disease. Vivek’s current focus is on developing scalable synthetic data generation strategies to solve challenging 3D computer vision problems in diagnostic and interventional radiology. More details: https://vivekg.dev/
Host Faculty: Prof. Phaneendra Yalavarthy
ALL ARE WELCOME
