Comparison of Pre-Trained vs Domain-Specific Convolutional Neural Networks for Classification of Interstitial Lung Disease

Interstitial Lung Disease (ILD) is an umbrella term used to describe different variations of lung diseases that affect humans. Due to the difficulty of classifying ILD, the time span taken to have a patients CT scans analyzed by a radiologist is rather long. To speed up the process, Computer-Aided Diagnostic (CAD) systems have been built. In this paper, one such approach based on Convolutional Neural Network (CNN) is proposed to classify ILD from CT scans. We investigate how a generalized pre-trained neural network compares to a domain-specific neural network. In addition, we propose different methods to help improve a CNNs performance in classifying ILD and assess how our data impacts performance for both models. The InceptionV3 model produced by Google Inc. yielded the best F1 score of 0.80 ± 0.11 rivaling a domain-specific model. When using an Ensemble Network composed of InceptionV3, VGG16, and ResNet50, the accuracy increased to 0.827 ± 0.08.