Convolutional Neural Network (CNN) for Interstitial Lung Disease (ILD) Patterns Recognition

The application of CNNs in medical image analysis has assumed a pivotal role, despite constraints related to the large amount of labeled data required. The aim of this study is to present a CNN for recognition and quantification of ILD patterns and to evaluate their correlations to functional data. 27 patients with chronic hypersensitivity pneumonitis (FEV1%p=61.6±15.9, FVC%p=57.0±16.8) were acquired via HRCT at full inspiration. The CNN (6 layers: 2 convolutional, 1 average pooling, 3 fully connected) was trained and tested on 395 regions of interest (ROIs) identified by one experienced pneumologist on the CT images resulting in a test accuracy=0.816, F1-score=0.810. The CNN was then applied to the patients’ full scans. 5 ILD patterns were analyzed: air trapping (AT), consolidation (C), ground glass opacity (GGO), healthy (H) and reticulation (R). The percentage of ILD tissue (ILD%) on the whole volume was correlated (Spearman Test) to spirometry. Fig. 1 reports CNN results on a representative patient. In the overall population, negative correlation was found between ILD% and FVC%p (r=-0.436, p<0.05), FEV1%p (r=-0.447, p<0.01) and TLC%p (r=-0.441, p<0.05). In conclusion, the proposed CNN allows the recognition of ILD with significant correlations to functional data. The classification of pathological ROIs into different ILD patterns may represent a first step towards the development of a computer aided diagnosis system.