There is considerable current interest in finding methods to improve the diagnostic accuracy of mammographic interpretation. We are studying one approach, in which the radiologist is presented with mammograms from past cases that the computer concludes are relevant to the case being evaluated, in effect providing an online atlas of images with known pathology. To make such an approach useful, it is important to define “relevance” of the images in an effective way. We were the first to propose the use of machine learning techniques to learn the concept of relevance directly from radiologists, in the form of similarity scores. We have shown that such an approach can yield to image retrieval results that closely match the preferences of the user.
An important component of today’s computer-aided diagnosis (CAD) toolkit is the ability to automatically detect clustered microcalcifications in digital mammograms. We have developed two modern machine-learning methods for calc detection—based on the relevance vector machine (RVM) and support vector machine (SVM)—which significantly outperform other techniques. This demonstrates that there is still a great deal of performance to be gained in CAD by introducing modern tools from statistical learning theory.
It is now widely accepted that the most appropriate way to judge image quality is to measure the ability of an image to serve its intended purpose. Toward this end, the channelized Hotelling observer (CHO) has been widely used to model human observer performance in medical lesion-detection tasks. We have developed an alternative approach, based on machine learning, in which we specifically aim to learn the human-observer model from actual data reported by humans in response to images presented to them. Specifically, we have found that support vector machines can significantly outperform the CHO in correctly modeling human observer performance. This work is sponsored by NIH/NHLBI.