Recent Advances in 3D Computer Vision and Their Applications for Medical Imaging
Building 9, Level 2, Hall1, Room 2325
Overview
Abstract
Advances in AI have revolutionized many domains, yet effectively harnessing 3D data remains a challenge due to the lack of a unified 3D representation across tasks. A recent breakthrough in this area is the 3D Gaussian Splatting primitive, which enables fast and accurate 3D reconstruction from multiple 2D images. However, its limitations in representing sharp, physical objects have prompted us to propose an alternative based on 3D smooth convexes, which demonstrably outperforms traditional approaches in reconstructing fine details.
In the realm of medical imaging—where 3D volumetric data is critical—conventional 3D understanding methods often adapt 2D vision models, potentially underutilizing rich 3D information. By reinterpreting 3D medical imaging as a video segmentation problem, our proposed MedSAM-2 model leverages video pretraining to achieve state-of-the-art segmentation performance on both 3D volumes and promptable 2D tasks. Complementing this approach, we introduce the UKBOB dataset, a large-scale collection of over one billion MRI segmentation labels covering full-body organs and bones, which serves as a foundation for a robust 3D segmentation model and sets new performance benchmarks in medical imaging.
Presenters
Abdullah Hamdi, Postdoctoral Research Fellow, Visual Geometry Group; Junior Research Fellow, Kellogg College, University of Oxford
Brief Biography
Abdullah is a postdoctoral research fellow in machine learning and computer vision at the Visual Geometry Group of the University of Oxford and a research felllow (JRF) at Kellogg College in Oxford University. Prior to that, he earned Ph.D. and MS degrees from KAUST, working on 3D understanding with deep neural networks, advised by Prof. Bernard Ghanem. Abdullah was also partly advised by Prof. Matthias Niessner at TUM in 2022 in Munich. Abdullah is a lead organizer of the 3DMV workshop at CVPR and is honored with multiple national and international distinctions. He is also the founder and president of fihm.ai, the largest Arabic online platform dedicated to teaching, educating, and spreading awareness about AI and deep learning technologies and applications.