I-DINet: A deep learning approach for multiclass classification of lung diseases in computer-aided diagnosis

Background Pneumonia and tuberculosis are both serious and life-threatening diseases, with pneumonia, caused by infections that inflame the air sacs in the lungs, and tuberculosis, caused by the bacterium Mycobacterium tuberculosis, affecting the lungs but can also spread to other parts of the body. Objective A deep learning (DL) approach to accurately diagnose and perform multi-class classification of lung diseases, namely pneumonia and tuberculosis, has been designed and implemented using chest computed tomography (CT) slices in this work. Methods The DenseNet-121 model is utilised to select the significant slices from the input dataset based on the provided labelled data. These selected slices have been augmented, and the augmented CT slices are further processed to utilise the U-Net model for segmenting the lung and lesion regions. The tuberculosis and pneumonia lesions considered in this work were cavitation, consolidation and ground-glass opacities (GGO). These lesions have been elucidated by an expert radiologist. A novel model, termed I-DINet (Improved DenseNet-121 and InceptionNet-V3), has been developed for feature extraction, and the extracted features have been used to train a densely connected layer within the network to perform classification. In this proposed I-DINet, selected layers from DenseNet-121 and Inception V3 networks have been retained, while others have been excluded. Furthermore, hyperparameter tuning has been conducted using a genetic algorithm (GA) to optimise the model's performance. Results The I-DINet model demonstrated superior results in terms of 98.76% accuracy, 98.87% recall, 99.38% specificity, 98.72% precision, and 98.79% F1 score when evaluated on a diverse public dataset compiled from various open-source platforms, including Kaggle, Mendeley, IEEE DataPort, and Radiopaedia. The I-DINet model has also been validated using open-source datasets from Kaggle (Montgomery and Shenzhen datasets) and a real-time dataset. Conclusion I-DINet outperforms standalone DenseNet-121 and Inception-V3 models, helping doctors make quicker and more precise diagnoses, which would ultimately result in improved patient outcomes.