Inferring the Performance of Medical Imaging Algorithms

摘要

Evaluation of the performance and limitations of medical imaging algorithms is essential to estimate their impact in social, economic or clinical aspects. However, validation of medical imaging techniques is a challenging task due to the variety of imaging and clinical problems involved, as well as, the difficulties for systematically extracting a reliable solely ground truth. Although specific validation protocols are reported in any medical imaging paper, there are still two major concerns: definition of standardized methodologies transversal to all problems and generalization of conclusions to the whole clinical data set. We claim that both issues would be fully solved if we had a statistical model relating ground truth and the output of computational imaging techniques. Such a statistical model could conclude to what extent the algorithm behaves like the ground truth from the analysis of a sampling of the validation data set. We present a statistical inference framework reporting the agreement and describing the relationship of two quantities. We show its transversality by applying it to validation of two different tasks: contour segmentation and landmark correspondence.