2016 Volume 34 Issue 1 Pages 38-42
Computer-Aided-Diagnosis (CAD) research and development is being conducted in diverse fields. However, these activities are hindered by the difficulty in acquiring case images. To deal with this, efforts are currently underway to artificially create case images by embedding tumors into lesion-free images. This approach adopts as its criteria the ability to produce artificial case images that are indistinguishable to a physician's eye from actual case images. Its usefulness in actual application to CAD, however, remains an unknown quantity. In this study, we developed liver tumor-extraction CAD and used artificial case images in clinical testing employing machine learning to investigate their usefulness. Also investigated was the impact on CAD performance when artificial and actual cases images were mixed. Leave-one-out cross-validation was used for investigation, by which assessment was conducted by switching between actual and artificial images at various fixed ratios. Artificial images bore comparison with actual images at proportions of 50% or less, suggesting the feasibility of artificial case images when used at fixed ratios.