Computer-aided detection (CADe) has been shown to increase readers? sensitivity and reduce inter-observer variance in detecting abnormalities in medical images. However, they prompt relatively large numbers of false positives (FPs) that readers find tedious to review and, during this process, the readers can incorrectly dismiss true lesions prompted correctly to them by CADe systems. Thus, there is a demand for an advanced decision support system that would provide not only high detection sensitivity, but also high specificity while being able to explain why a specific location was prompted as a lesion. In this project, we propose to improve the detection specificity of CADe by deep convolutional neural networks (DCNNs) that can analyze the extrinsic radiomic phenotype, such as the context of local anatomy, of target lesions, whereas current CADe systems consider only the intrinsic radiomic phenotype, such as the shape and texture of detected lesions. Further, we can use DCNNs to provide an explanation of why a specific location was prompted by using anatomically meaningful object categories with similar-image retrieval of past diagnosed cases. In this project, we will focus on computed tomographic colonography (CTC), which is a minimally invasive screening method for early detection of colorectal lesions to prevent colorectal cancer (CRC), which is the second leading cause of cancer deaths in the United States. Historically, however, only adenomas were believed to be precursors of CRC. Recent studies have revealed a molecular pathway where also serrated lesions can develop into CRC. Recent studies have indicated that CTC can detect serrated lesions accurately based upon the phenomenon called contrast coating. Thus, the goal of this project is to develop a deep radiomic decision support (DeepDES) system that leverages deep learning for providing high sensitivity and specificity in the detection of colorectal lesions, in particular, serrated lesions, and for providing diagnostic information that explains why a specific location was prompted as a lesion to assist readers in assessing detected lesions correctly. To achieve the goal, we will explore the following specific aims: (1) Develop a radiomic deep-learning (RAID) scheme for the detection of colorectal lesions, (2) develop a DeepDES system for diagnosis of detected lesions, and (3) evaluate the clinical benefit of DeepDES system. Successful development of the proposed DeepDES system will provide an advanced decision support that addresses the current concerns about CADe by yielding both high detection sensitivity and high specificity while being able to explain why a specific location was prompted as a target lesion. Broad adoption and use of the DeepDES system will advance the prevention and early diagnosis of cancer, and thus will ultimately reduce mortality from colorectal cancer in the United States.
Successful development of the proposed deep radiomic decision support (DeepDES) system will provide an advanced decision support that addresses the current concerns about computer-aided detection (CADe) by yielding both high detection sensitivity and high specificity while being able to explain why a specific location was prompted as a target lesion. Such a system is expected to outperform current state-of-the-art CADe systems in the diagnosis of colorectal lesions, in particular, serrated lesions. Broad adoption and use of the DeepDES system will advance the prevention and early diagnosis of cancer, and thus will ultimately reduce mortality from colorectal cancer in the United States.
