Studies on computer‐aided diagnosis systems for chest radiographs and mammograms (in Japanese)

Abstract

This thesis describes computer‐aided diagnosis (CAD) systems for chest radiographs and mammograms. Preprocessing and imaging processing methods for each CAD system include dynamic range compression and region segmentation technique. A new pattern recognition technique combines genetic algorithms with template matching methods to detect lung nodules. A genetic algorithm was employed to select the optimal shape of simulated nodular shadows to be compared with real lesions on digitized chest images. Detection performance was evaluated using 332 chest radiographs from the database of the Japanese Society of Radiological Technology. Our average true‐positive rate was 72.8% with an average of 11 false‐positive findings per image. A new detection method using high resolution digital images with 0.05 mm sampling is also proposed for the mammogram CAD system to detect very small microcalcifications. An automated classification method uses feature extraction based on fractal dimension analysis of masses. Using over 200 cases to evaluate the detection of mammographic masses and calcifications, the detection rate of masses and microcalcifications were 87% and 96% with 1.5 and 1.8 false‐positive findings, respectively. The classification performance on benign vs malignant lesions, the Az values that were defined by the areas under the ROC curves derived from classification schemes of masses and microcalcifications were 0.84 and 0.89. To demonstrate the practicality of these CAD systems in a computer‐network environment, we propose to use the mammogram CAD system via the Internet and WWW. A common gateway interface and server‐client approach for the CAD system via the Internet will permit display of the CAD results on ordinary computers.