报告时间：2025年3月28日（周五）15:30-16:30
报告地点：东南大学九龙湖校区计算机楼513报告厅
报告人：程雪莲 研究员（兼讲师） 蒙纳士大学
报告摘要: Recent advances in 3D reconstruction have demonstrated promising improvements in training efficiency and reconstruction quality. However, significant challenges persist in medical imaging applications. This report explores two key applications of 3D reconstruction techniques: Robotic-Assisted Minimally Invasive Surgery (RAMIS) and Cone-Beam Computed Tomography (CBCT).
The first focus area introduces endoscopic reconstruction for RAMIS. Compared to traditional 2D monitoring, 3D reconstruction offers significant advantages by enabling users to observe the surgical site from any angle. Our goal is to generate3D models of observed tissues from stereo endoscopic video feeds. The reconstruction pipeline employs stereo vision matching algorithms to estimate depth maps, followed by the fusion of RGBD images into 3D models.
The second focus area addresses sparse-view reconstruction in CBCT imaging. Given the potential radiation risks associated with X-ray imaging in CT scans, sparse-view reconstruction CBCT emerges as a promising solution by maintaining imaging quality while utilizing only a few dozen projections. Our investigation explores the application of multiview reconstruction principles to sparse-view CBCT. We present a novel framework that combines diffusion models with 3D Gaussian Splatting for sparse-view CBCT reconstruction, effectively reducing artifacts and hallucinations.
Through comprehensive case studies and practical examples, this report demonstrates the potential integrating these advanced 3D reconstruction techniques in medical applications. This capability provides substantial benefits for downstream tasks, e.g. surgical navigation and surgeon-centric augmented and virtual reality implementations.
报告人简介：Dr. Xuelian Cheng currently is a Research Fellow at Monash Suzhou and an Adjunct Lecturer at Monash University Australia. Prior to this role, She completed PhD at Monash University under the guidance of A/Prof. Zongyuan Ge , A/Prof. Mehrtash Harandi , and Prof. Tom Drummond . She had contributed to diverse research projects with industrial companies including Tencent (Canberra XR lab), IIAI, Airdoc and SenseTime. The outcomes resulted in publications in top ML/CV and medical imaging conferences. She was a research intern at the Mobile Intelligence Group (MIG) of SenseTime Technology in Beijing, China (July 2019 to November 2019). She had the opportunity to work remotely as a research intern at IIAI and work with scholars at MBZUAI (April 2021 to October 2021). She worked as a research intern at Tencent XR Lab in Canberra, Australia (August 2022 to February 2023).
Previous representative works include deep learning for 3D visual perception and reconstruction, automated machine learning (AutoML), video analysis and object detection. With a keen interest in the interdisciplinary realm of machine learning, computer vision, medical image analysis, and robotic surgery intelligence, she aspires to leverage her expertise in ML/CV to pioneer novel approaches for disease detection, diagnosis, and monitoring, as well as surgical interventions. Her focus extends to the application of visual medical data (AR/VR/XR) technologies for educational purposes, aiming to make significant strides in enhancing patient outcomes and revolutionizing medical practice.
Her representative research includes a novel solution for learning effective architectures for deep stereo matching (LEAStereo). This solution ranked first on KITTI stereo and Middlebury benchmarks for half a year, which has outperformed all existing state-of-the-art deep stereo matching architectures. This work has also provided insights into the following works and has potential commercial applications. All the publications can be evidenced by checking easily via her Google Scholar.
