Advancing Interstitial Lung Disease Diagnosis: a CNN Approach for High-Resolution Computed Tomography Image Classification

Interstitial Lung Diseases (ILDs) are an important heterogeneous group of rare diseases that are that difficult to predict, diagnose and monitor for clinical signs of low specificity and radiological signs that require manual radiologist work. Data-driven and advanced image-processing support for ILD diagnostics is emerging rapidly. We designed a Convolution Neural Network (CNN) for the classification of five classes of ILD patterns in High-Resolution Computed Tomography (HRCT) images. Our CNN consists of three convolutional blocks and one flatten layer, followed by three dense layers. The activation function of the first two dense layers is ReLU, the last dense layer uses SoftMax activation function to provide probability values at the output for the five classes. The proposed CNN was trained and evaluated on the publicly available ILD dataset of University Hospitals of Geneva (HUG). Our simple model showed high performance, achieving an average F1 score of 0.86 across classes, with scores above 0.96 in some of them.