The US Surgeon General has declared pulmonary embolism (PE) a major national health problem, causing more deaths than breast, colon, and lung cancers. The current diagnostic standard for suspected PE is CT pulmonary angiography (CTPA). However, the number of CTPA examinations is increasing dramatically, and incorrect CTPA interpretations are frequent in general practice (10-14% over/under-diagnosis). There is a clinical need to improve the efficiency and accuracy of PE diagnosis at CTPA. Our central hypothesis is that this clinical need can be addressed by exploiting computer-radiologist synergy. However, existing computer-aided diagnosis (CAD) methods for PE have serious deficiencies: they are limited in sensitivity and specificity, incapable of handling PE over-diagnosis, and operating only at the embolus level?localizing individual emboli, but PE diagnosis is rendered at the patient-level?excluding non-PE patients and dispatching PE-patients to treatment. Therefore, our objective is to overcome these deficiencies with a new methodology. We have built a strong interdisciplinary team, developed an innovative prototype, and evaluated it through our pilot clinical studies, demonstrating outstanding performance. This proposed research has three specific aims: 1) boost our current system?s embolus-level performance with our newly proposed strategies, assisting radiologists in accurately localizing emboli and facilitating precision medicine through risk stratification; 2) achieve patient-level diagnosis through our newly developed algorithms, assisting radiologists in quickly excluding negative patients and improving diagnostic efficiency; and 3) demonstrate clinical benefits of our system by testing specific clinical hypotheses. This research is innovative because (1) our approach to embolus-level detection fundamentally differs from prior approaches in that it requires no vessel segmentation, overcoming their limitations; (2) we are pioneering two uncharted areas: PE patient-level diagnosis and over-diagnosis prevention; we do not perceive any similar objectives in existing NIH grants or publications in the literature; and (3) this project utilizes our original algorithms and will yield multiple novel algorithms. Our project is significant because it (1) addresses a major national health problem; (2) develops a new methodology that transcends the current paradigm from mere detection of emboli to simultaneous patient-level diagnosis, embolus-level detection, and over-diagnosis prevention, overcoming the deficiencies of the current PE CAD systems; and (3) delivers a next- generation, high-performance PE CAD system that quickly excludes non-PE patients, accurately localizes emboli, and actively prevent PE over-diagnosis, thereby enhancing radiologists? diagnostic capabilities and supporting precision medicine through risk stratification. Successful completion of the project is expected because (1) we have already made good progress in algorithm development and clinical evaluation; (2) our approach is carefully crafted on solid algorithmic and mathematical foundations; (3) our clinical evaluation is rigorously designed; and (4) our team is uniquely capable and well prepared to conduct this project, which builds upon our innovative research in CAD, pioneering research in deformable models, and world-renowned PIOPED trials. This research is expected to have important impact on PE- related clinical practice, development of decision support systems for many diseases, and medical education.
Behind the great success of biomedical imaging, a crisis is looming: the number of imaging studies is growing exponentially; the workload of radiologists is increasing dramatically; the health-care cost related to imaging is rising rapidly?We are facing a grant new challenge: ?image data explosion? (a manifestation of big data in biomedical imaging): Modern imaging systems generate enormous data, far exceeding human abilities for interpretation, but what is paramount are not the images themselves, rather the clinically relevant information contained within the images; therefore, our long-term goal is to develop and validate comprehensive, high-performance computational tools that automatically and quantitatively extract clinically relevant information from images to support clinical decision making and facilitate precision medicine. To demonstrate the immediate, measurable impact of our research, we have chosen pulmonary embolism as our initial research platform because the Surgeon General has declared pulmonary embolism a major national health problem, causing more deaths than breast cancer, AIDS, and motor vehicle accidents combined; with laudable efforts to diagnose pulmonary embolism, the number of CT studies for suspected pulmonary embolism has been increasing dramatically, while incorrect CT interpretations are frequent in general practice (10-14% over/under-diagnosis); therefore, there is a clinical need to (1) mitigate radiologists? workloads and (2) improve the efficiency and accuracy for pulmonary embolism diagnosis using CT. Our objective is to address this clinical need by exploiting radiologist-computer synergy, delivering two important outcomes: (a) a new methodology that transcends the current paradigm from mere detection of emboli to simultaneous patient-level diagnosis, embolus-level detection, and over-diagnosis prevention; and (b) a next-generation, high- performance system that will be able to assist radiologists in quickly excluding negative patients without overlooking positive patients, accurately localizing individual emboli to support personalized treatments through risk stratification, and actively preventing PE over-diagnosis, exerting an important positive impact on clinical practice associated with pulmonary embolism?a life-threatening condition.
| Zhou, Zongwei; Shin, Jae; Zhang, Lei et al. (2017) Fine-tuning Convolutional Neural Networks for Biomedical Image Analysis: Actively and Incrementally. Proc IEEE Comput Soc Conf Comput Vis Pattern Recognit 2017:4761-4772 |
