Database Acquisition for the Lung Cancer Computer Aided Diagnostic Systems

Abstract

Most of the used computer aided diagnostic (CAD) systems based on applying the deep learning algorithms are similar from the point of view of data processing stages. The main typical stages are the training data acquisition, pre-processing, segmentation and classification. Homogeneity of a training dataset structure and its completeness are very important for minimizing inaccuracies in the development of the CAD systems. The main difficulties in the medical training data acquisition are concerned with their heterogeneity and incompleteness. Another problem is a lack of a sufficient large amount of data for training deep neural networks which are a basis of the CAD systems. In order to overcome these problems in the lung cancer CAD systems, a new methodology of the dataset acquisition is proposed by using as an example the database called LIRA which has been applied to training the intellectual lung cancer CAD system called by Dr. AIzimov. One of the important peculiarities of the dataset LIRA is the morphological confirmation of diseases. Another peculiarity is taking into account and including “atypical” cases from the point of view of radiographic features. The database development is carried out in the interdisciplinary collaboration of radiologists and data scientists developing the CAD system.