Application of deep learning neural network for classification of TB lung CT images based on patches

In this work, convolutional neural network (CNN) is applied to classify the five types of Tuberculosis (TB) lung CT images. In doing so, each image has been segmented into rectangular patches with side width and high varying between 20 and 55 pixels, which are later normalised into 30x30 pixels. While classifying TB types, six instead of five categories are distinguished. Group 6 houses those patches/segments that are common to most of the other types, or background. In this way, while each 3D dataset only has less than 10% distinguishable volumes that are applied to perform the training, the rest remains part of the learning cycle by participating to the classification, leading to an automated process to differentiation of five types of TB. When tested against 300 datasets, the Kappa value is 0.2187, ranking 5 among 23 submissions. However, the accuracy value of ACC is 0.4067, the highest in this competition of classification of TB types.

Paper published as: X. Gao, Y. Qian, Application of Deep Learning Neural Network for Classification of TB lung CT Images based on Patches. CLEF 2017 Working Notes. Working Notes of CLEF 2017 - Conference and Labs of the Evaluation Forum
Dublin, Ireland, September 11-14, 2017. Linda Cappellato, Nicola Ferro, Lorraine Goeuriot, Thomas Mandl (eds), CEUR-WS.org, online: http://ceur-ws.org/Vol-1866/paper_87.pdf