J-Net: An Adaptive System for Computer-Aided Diagnosis in Lung Nodule Characterization

The aim of this study is to evaluate the capability of improved artificial adaptive systems and additional novel training methods in order to distinguish between benign and malignant lung nodules in Multi Detector Computed Tomography. A total of 90 nodules belonging to 88 patients are analyzed. A set of adjacent slices representing the lesion selected from the CT Image analysis by the experts are collected and stored in a database. Features extracted by an assembly of adaptive algorithms working in sequence [Active Connection Fusion (ACF): a new set of ANNs for image fusion; J-Net Active Connections Matrix (J-Net): a new ANN for dynamic image segmentation; Population (Pop): a new and fast multidimensional scaling algorithm] are divided into several groups using random or experimental methods to train and test different Artificial Neural Networks. Best results are obtained with Adaptive Learning Quantization (AVQ) and Meta-Consensus, two new supervised ANNs, experts in rapid classification and not sensitive to over fitting. After optimization of the distribution of cases among the training and testing sets the following results are achieved: sensitivity (recognition of malignant nodules) ranging from 93.33% to 100%; specificity (recognition of benign nodules) stable on 95.56% and overall accuracy ranging from 94.44% to 97.78%. These results represent the highest predictive values ever recorded in lung CAD literature. Benchmarking analysis, with advanced mathematical algorithms using simpler approaches, show that complex processing systems, composed of different steps and sub processing systems, are clearly superior and probably needed to reach excellent predictive performances in lung nodule characterization.