Lung Hrct Pattern Classification for Cystic Fibrosis Using Convolutional Neural Network Publisher



Nezamabadi K¹ ; Naseri Z¹ ; Moghaddam HA¹ ; Modarresi M^{2, 3} ; Pak N⁴ ; Mahdizade M⁴
Source: Signal# Image and Video Processing Published:2019

Abstract

Cystic fibrosis (CF) is one of the most prevalent autosomal recessive disorders among whites. It mostly affects lungs which causes infection and inflammation leading to 90% of deaths among CF patients. Due to wide variability in its clinical presentation and organ involvement, studying responses to therapy and evaluation of pulmonary changes over time is crucial in progress prevention of CF. Serial high-resolution computed tomography (HRCT) scans significantly facilitate the assessment of the pulmonary abnormalities evolution in CF patients. Recently, artificial intelligence is being employed for analyzing thoracic CT scans acquired from CF patients. In this paper, we propose a convolutional neural network (CNN) approach for classifying CF lung patterns in HRCT images. The proposed network consists of two convolutional layers with 3 × 3 kernels and max-pooling in each layer followed by two dense layers of 1024 and 10 neurons, respectively. The softmax layer prepares a probabilistic distribution of predicted output among classes. This layer has three outputs, equivalent to the classes corresponding to normal (healthy), bronchiectasis, and inflammation. To train and evaluate the network, we built up a patch-based dataset extracted from more than 1100 lung HRCT slices which were acquired from 45 CF patients. A comparative evaluation proved the effectiveness of the proposed CNN with respect to its close counterparts. The classification accuracy, mean sensitivity, and specificity of 93.64%, 93.47%, and 96.61% were achieved which demonstrated the potential of CNNs in analyzing lung CF patterns and following up the lungs’ status. In addition, the visual features extracted by our proposed method might be helpful for automatic measurement and eventually scoring the severity of CF patterns in lung HRCT images. © 2019, Springer-Verlag London Ltd., part of Springer Nature.