Abstract
Aiming to compensate for the lack of case images in computer-aided diagnosis/detection (CAD) development, efforts have been made to artificially create case images by embedding lesions (such as tumors) into lesion-free images. We provide some examples of the usefulness of such images in liver tumor CAD, verified by judgment equipment that was designed using training data including artificial cases. For this study, we created an artificial case database (DB) made up of different variations, in consideration of practical use, proposing combinations and methods for generating optimal performance enhancement to unknown data in CAD. We created learning data for combinations that emphasized a lesion as the artificial case images with a variety of sizes and contrast changes, and we investigated methods of constructing optimal learning data, based on performance of the judgment equipment that we designed against unknown data, and we confirmed its effectiveness. In order to further confirm the generalizability of this technique with regard to application to other regions, we also confirmed that this is effective when applied to breast cancer tumor CAD. Additionally, we also propose using an artificial case DB as a performance evaluation measure for the various types of CAD. That is, for lesions (such as tumors) of standard size and contrast, we believe we can obtain objective evaluation criteria by comparing detection performance in response to quantitative changes in size and brightness.
