REDUCTION OF FALSE POSITIVES IN LUNG NODULE DETECTION USING A 2-LEVELNEURAL CLASSIFICATION

We have developed a neural-digital computer-aided diagnosis system, ba sed on a parameterized two-level convolution neural network (CNN) arch itecture and on a special multilabel output encoding procedure. The de veloped architecture was trained, tested, and evaluated specifically o n the problem of diagnosis of lung cancer nodules found on digitized c hest radiographs. The system performs automatic ''suspect'' localizati on, feature extraction, and diagnosis of a particular pattern-class ai med at a high degree of ''true-positive fraction'' detection and low ' 'false-positive fraction'' detection. In this paper, we aim at the pre sentation of the two-level neural classification method in reducing fa lse-positives in our system. We employed receiver operating characteri stics (ROC) method with the area under the ROC curve (A(z)) as the per formance index to evaluate all the simulation results. The two-level C NN showed superior performance (A(z) = 0.93) to the single-level CNN ( A(z) = 0.85). The proposed two-level CNN architecture is proven to be promising and to be extensible, problem-independent, and therefore, ap plicable to other medical or difficult diagnostic tasks in two-dimensi onal (2-D) image environments.