Real-Time Bronchoscope Localization Using Machine Learning To Improve Lung Cancer Diagnosis

点击翻译按钮获取中文摘要

Project Summary/Abstract Lung cancer 5-year survival rates drop from 61% for early stage diagnosis to just 6% for late stage diagnosis. Currently, fewer than 1 in 5 cases are diagnosed at an early stage. The increasing frequency of chest CT scans and changes in lung cancer screening guidelines are expected to increase the number of incidentally discovered lung lesions, representing an opportunity for earlier lung cancer diagnosis. Bronchoscopy is currently the safest, least invasive, and least expensive diagnostic option, but its poor diagnostic yield greatly limits its procedural benefit. Even when advanced techniques like radial endobronchial ultrasound and electromagnetic navigation are used, the diagnostic yield is just 50-60%. This is primarily due to challenges with intraoperative localization of the bronchoscope prior to needle deployment. Additionally, access to these techniques is limited because they require expensive equipment and unique expertise. Efforts relying on the bronchoscope's built-in camera require no additional equipment or specialization, but have struggled with generalizability across individuals in part due to limited data availability and assumptions about airway features. The objective of this proposal is to improve the success rate of traditional bronchoscopes by addressing limita- tions in intraoperative localization using a data-driven model that is robust to differences in human anatomy. This work has potential for significant public health benefit by (1) increasing early lung cancer detection, (2) reducing morbidity and mortality by reducing the number of invasive procedures, and (3) making minimally invasive bron- choscopy more accessible in areas without expert bronchoscopists. The proposed work will be accomplished via two Specific Aims. In Aim 1, a dataset will be generated of virtual and real bronchoscopy videos with video-frame matched six degrees-of-freedom poses (position and orientation in three-dimensions) of the bronchoscope's dis- tal tip. This data will be made publicly available as the first large dataset of its kind to promote future research and reproducibility. In Aim 2, a real-time bronchoscope localization model will be developed using advances in ma- chine learning, including deep neural networks, that have shown success in camera localization for non-medical applications. These models will regress the pose of the bronchoscope's distal tip using current and past video frames of the bronchoscope's built-in camera. The clinical utility of the system will be evaluated in simulation, 3D printed lung phantoms, and ex-vivo porcine lung experiments. The research, tightly coupled clinical experience, and associated training plan will provide a unique interdisciplinary skill-set in computer science, medical robotics, and procedural medicine. The outstanding research and clinical environment for this training at the University of North Carolina at Chapel Hill ensures exceptional preparation for a career conducting cutting-edge research as a physician-scientist in medical robotics.