Lung image Ssgmentation using deep learning methods and convolutional neural networks

Abstract

This paper presents results of the first,
\nexploratory stage of research and developments on
\nsegmentation of lungs in X-Ray chest images (Chest
\nRadiographs) using Deep Learning methods and
\nEncoder-Decoder Convolutional Neural Networks (EDCNN).
\nComputational experiments were conducted using
\nGPU Nvidia TITAN X equipped with 3072 CUDA Cores
\nand 12Gb of GDDR5 memory. Comparison of resultant
\nsegmentation accuracy with manual segmentation using
\nDice's score has revealed that the average accuracy
\nachieves 0.962 with the minimum and maximum Dice's
\nscore values of 0.926, 0.974 respectively, and standard
\ndeviation of 0.008. The study was performed in the
\ncontext of large-scale screening of population for lung
\nand heart diseases as well as development of
\ncomputational services for international portal on lung
\ntuberculosis. The results obtained with this study allow
\nconcluding that ED-CNN networks may be considered as
\na promising tool for automatic lung segmentation in
\nlarge-scale projects.