Lung Disease Detection Using CNNs and Transfer Learning on Medical Imaging

Lung disease are some of life's most prevalent lung illnesses, which significantly affect morbidity and mortality. The main reason to diagnose these diseases is to ensure early intervention and appropriate quality of treatment to the patients. Owing to the datasets, this work propounds an automated solution for a multilabel lung disease classification based on the CNN and Transfer Learning empowered by the NIH Chest Xray Dataset. Accompanying the competition is the dataset of 112,120 chest X-ray images across 14 disease categories that provide a reliable ground for training and testing. The ResNet and DenseNet models were fine-tuned, and multiple diseases were classified, and class imbalance of the diseases was addressed through augmentation and weighted losses. The proposed model yielded an accuracy of $\mathbf{9 4 \%}$ as well as precision, recall, and F1 score higher than the recognized models such as CheXNet and traditional Machine Learning methods, including Random Forest and SVM. We also extended Grad-CAM for interpretability, which visually explained the important regions that affected the predictions, thus improving clinical relevance. The overall performance for common diseases was satisfying, although identification of especially rare forms of diseases such as pneumothorax and fibrosis proved difficult and carefully optimized through the Use of disease-specific measures. With CNNs and Transfer Learning, this study shows that traditional medical diagnostics could be revolutionized with scalable and efficient solutions that doctors can apply. Further work involves incorporating multiple modalities of clinical data and CT scans, as well as fine tuning the model for practical applications. These developments are intended to design a friendly, stable, and explainable system to help radiologists, especially in limited-resource environments.