Learning the visual and cognitive bases of lung nodule detection

项目总结/文摘

Project Summary/Abstract Lung cancer is the most frequent cause of cancer death in the United States among both men and women. If lung nodules can be detected with greater reliability at an early stage, significant improvements in survival rate would be achievable. Chest radiographs are among the most common diagnostic tool used in radiology, and can reveal unexpected incidences of lung cancer. However, even expert radiologists may fail to detect the presence of a subtle low-contrast pulmonary nodule against the high-contrast anatomical background of a chest X-ray, with estimated rates of missed detection of 20-30%. What are the perceptual mechanisms, cognitive mechanisms, and critical learning experiences that determine how well a person can perform this challenging task of lung nodule detection? The PI and Co-Investigator have formed a synergistic collaboration that brings together expertise in human vision, computational modeling and neuroscience (Dr. Tong) in concert with thoracic imaging and biomedical engineering (Dr. Donnelly) to address this longstanding problem with high clinical relevance. This project will develop a validated computational approach for generating a diverse set of visually realistic simulated nodules to achieve the following goals. These are: 1) to characterize radiologist performance on an image-by-image basis in an ecologically valid manner, 2) to develop a novel image- computable model that accounts for expert performance, and 3) to develop a novel learning-based paradigm to characterize the perceptual and cognitive mechanisms of nodule detection, initially in non-expert participants, with the long-term goal of developing a protocol to enhance clinical training. The project will incorporate sophisticated 2D image-based computational methods as well as data from 3D CT segmented nodules to generate a diverse set of simulated nodule examples, each placed in a unique chest X-ray. Success will be evaluated by the following outcome measures. First, radiologists should find it very difficult to tell apart real from simulated nodules. Moreover, their performance accuracy at detecting/localizing simulated nodules should be predictive of their accuracy for real nodules. Second, if the simulated nodules suitably capture the variations of real nodule appearance, then non-expert participants who receive multiple sessions of training with simulated nodules should show improved performance for both simulated and real nodules. This learning- based paradigm will allow for characterization of the perceptual, cognitive, and learning-based factors that govern nodule detection performance. Third, development and refinement of this learning-based paradigm should have the potential to improve nodule detection performance in radiology residents. Finally, the behavioral data gathered from radiologists and other top-performing participants will be used to develop an image-computable model of nodule detection performance. As a whole, this project will lead to a more rigorous understanding of the perceptual and cognitive bases of lung nodule detection, and spur the development of a new learning-based protocol to enhance the training of radiology residents and other medical professionals.