A hybrid classifier for automated radiologic diagnosis: preliminary results and clinical applications

We describe the design, implementation, and preliminary evaluation of a computer system to aid clinicians in the interpretation of cranial magnetic-resonance (MR) images. The system classifies normal and pathologic tissues in a test set of MR scans with high accuracy. It also provides a simple, rapid means whereby an unassisted expert may reliably label an image with his best judgment of its histologic composition, yielding a gold-standard image; this step facilitates objective evaluation of classifier performance. This system consists of a preprocessing module; a semiautomatic, reliable procedure for obtaining objective estimates of an expert's opinion of an image's tissue composition; a classification module based on a combination of the maximum-likelihood (ML) classifier and the isodata unsupervised-clustering algorithm; and an evaluation module based on confusion-matrix generation. The algorithms for classifier evaluation and gold-standard acquisition are advances over previous methods. Furthermore, the combination of a clustering algorithm and a statistical classifier provides advantages not found in systems using either method alone.