FibNet: A Deep Learning-Based Interpretable System for Automated Diagnosis of Pulmonary Fibrosis

Pulmonary fibrosis (PF) is a progressive interstitial lung disease (ILD) that severely impairs respiratory function and can lead to fatal outcomes. Accurate and early diagnosis from chest computed tomography (CT) scans is essential for timely treatment but remains challenging due to visual similarities with other lung conditions. Conventional manual assessment is time-consuming and prone to inter-observer variability, while traditional deep learning models like VGG16, ResNet50, and DenseNet121 often struggle with optimal accuracy and clinical interpretability. To address these challenges, we propose FibNet, an EfficientNet-B3-based deep learning framework enhanced with Squeeze-and-Excitation (SE) attention and Grad- CAM++ for improved accuracy and explainability. The model incorporates preprocessing steps including lung segmentation and Hounsfield unit normalization to enhance feature representation. Validation was conducted on two distinct publicly available datasets: the Open Source Imaging Consortium (OSIC) PF Progression dataset comprising 176 patients with more than 24,000 CT slices, and the ILD dataset from the Lung Tissue Research Consortium containing 128 patients with approximately 18,000 CT slices. Each dataset was split 70:30 for training and testing. The implementation was carried out using TensorFlow 2.12 with the Keras API on an NVIDIA RTX 3090 GPU (24 GB VRAM). Experimental results show that FibNet achieved an accuracy of 0.978, Dice coefficient of 0.97, IoU of 0.95, and Matthews Correlation Coefficient (MCC) of 0.96 on the OSIC dataset, and 0.968, 0.96, 0.93, and 0.95 respectively on the ILD dataset, outperforming all baseline models. In conclusion, FibNet provides a computationally efficient and interpretable solution for PF diagnosis, with robust validation across multiple datasets, offering strong potential for real-world clinical integration.