A new approach to differential lung diagnosis with CT scans based on the Siamese neural network
Анна Александровна Мелдо, Lev Vladimirovich Utkin · Journal of Physics Conference Series · 2019
Abstract A lot of computer-aided diagnosis systems for lung cancer detection have been developed in the last years, but most of them may be not effective when we deal with the differential diagnosis. Only a part of patients with lung cancer have a typical nodular cancer. Some patients have the extremely difficultly recognized lung cancer due to its atypical visualization. A new approach is proposed for differential lung diagnosis and for solving the problem of classifying all types of lung cancer. It is based on the following ideas. First, we apply the length chord and Hounsfield unit value histograms characterizing the inner structure of a tissue and its surrounding as a new feature representation of the tissue. Second, we collect a special dataset which contains a lot of atypical cases of cancer and non-cancer tissues. Third, we propose to use histograms as well as suspicious lung object images for training and using two Siamese neural networks which can be viewed as a key element of the proposed approach and allow us to implement some elements of explainable artificial intelligence systems.