Deep Learning-Based Medical Image Analysis for Automated Disease Detection and Diagnosis
TL;DR
Results showed that the proposed DenseNet121-Attention framework is effective in efficiently and interpretively classifying the chest X-rays and can be used to support computer-aided diagnosis (CAD).
Abstract
In the medical field, automated medical image analysis is a key application of deep learning that aids in fast and uniform diagnosis of a disease. In this study, we propose DenseNet121-Attention framework for three class chest X-ray classification: Normal, Bacterial Pneumonia and Viral Pneumonia. The method proposed involves image preprocessing, data augmentation, transfer learning-based deep feature extraction, attention-driven feature refinement, Softmax classification and Grad-CAM based interpretability. DenseNet121 is used to extract hierarchy of radiological features and the attention mechanism is used to focus on the pulmonary areas of importance and to ignore less relevant information from the image. The overall accuracy, precision, sensitivity, F1-score and macro-AUC achieved are 84.29%, 84%, 83%, 83% and 94.54%, respectively. The class-wise analysis reveals excellent detection of bacterial pneumonia (94% sensitivity), high precision of the Normal class (93%) and high AUC score (96.10%). Results showed that the proposed framework is effective in efficiently and interpretively classifying the chest X-rays and can be used to support computer-aided diagnosis (CAD). Future directions will involve working with more multicenter datasets, multimodal imaging approaches, working with a more transformer-like model, and enhanced explainable AI methods.