Attention focusing front-end to detect lung nodules in chest radiograms

The paper describes an attention focusing system for computer-aided detection of lung nodules in chest radiograms. The main design requirement is the achievement of a high sensitivity along with adequate specificity. Nodular lesions are often subtle and extremely variable structures embodied in a complex background. Our approach is modular and based on multi-scale processing. Salient image features are enhanced by biologically inspired filters (both LoG and Gabor kernels). System architecture includes artificial neural networks of the feed-forward type which allows an efficient use of a priori knowledge about nodule shape and background structure. To test the achievable performances we have built an annotated radiogram data-set including 150 images with 233 (benign and malignant) nodules. A subset, including 100 images, was used to train the system while the remaining 50 images were used for testing purposes. Experimental results are summarized by ROC analysis.