An Ensemble of Triplet Neural Networks for Differential Diagnostics of Lung Cancer

A new classification subsystem of a lung cancer computer-aided-diagnosis systems is proposed in the paper. Its implementation is based on two main approaches. First, the computed tomography images of segmented suspicious lung nodules are represented by means of five histograms characterizing the shape, inner and outer structures of nodules. This representation significantly reduces the dimensionality of data. Second, an ensemble of triplet neural networks is used to take into account atypical cases of lung cancer and to improve accuracy of the classification subsystem usage. An architecture of the developed triplet network and peculiarities of the triplet network ensemble training process are considered in detail. The corresponding results of numerical experiments with using public dataset LUNA16 show outperforming properties of the proposed classification subsystem.