An X-ray CAD system with ribcage suppression for improved detection of lung lesions

Abstract

The purpose of our study is to prove that eliminating bone shadows from chest radiographs can greatly improve the accuracy of automated lesion detection. To free images from rib and clavicle shadows, they are first segmented using a dynamic programming approach. The segmented shadows are eliminated in difference space. The cleaned images are processed by a hybrid lesion detector based on gradient convergence, contrast and intensity statistics. False findings are eliminated by a Support Vector Machine. Our method can eliminate approximately 80% of bone shadows (84% for posterior part) with an average segmentation error of 1 mm. With shadow removal the number of false findings dropped from 2.94 to 1.23 at 63% of sensitivity for cancerous tumors. The output of the improved system showed much less dependence on bone shadows. Our findings show that putting emphasis on bone shadow elimination can lead to great benefits for computer aided detection.