Comprehensive Experimentation of Pretrained Models on Slice-Based Classification of Interstitial Lung Disease Patterns

Abstract

Interstitial Lung Diseases (ILD) are typically progressive diseases characterized by poor prognosis due to the inflammation and fibrosis affecting lung tissue. ILD is diagnosed through the identification of specific patterns or combinations of patterns that occur in various regions of the lung. This study employs High-Resolution Computed Tomography (HRCT) scans from the MedGIFT database to classify the patterns causing ILD on a slice-based. To achieve this, the pretrained models and a base Convolutional Neural Network (CNN) are utilized to provide a slice-based classification of ILD patterns in five, six, and seven classes. Four different pretrained models, namely VGG, DenseNet, MobileNet, and EfficientNet, are employed, and the performance impact of two training strategies, namely transfer learning and fine-tuning, is also evaluated. In the study, the effects of four different input resolution types on classification performance were investigated. The features extracted from the pretrained models and a base CNN are classified using a fully connected Artificial Neural Network classifier. The classification performance was further examined using two data augmentation methods for the most successful model and input resolution types. With the EfficientNetB0 pretrained model, classification results of five, six, and seven classes are obtained as 98.070%, 90.819%, and 87.781% F-score, respectively. Additionally, the computational costs and time complexity of all model combinations are analyzed, and their characteristics are comparatively discussed. This study contributes to the limited body of research on slice-based classification and advances clinical practice by facilitating the automatic detection of patterns on HRCT slices as a preprocessing step. Furthermore, the MedGIFT database is systematically analyzed in terms of slice and Region of Interest numbers across different pattern types, offering meaningful insights to support and guide its use in future research.