DEEP LEARNING APPROACHES FOR MEDICAL IMAGING ANALYSIS: TRANSFORMING DISEASE DETECTION AND MONITORING

Abstract

Background: Deep learning has been a revolution in the medical imaging sector, as it has offered improved accuracy and reproducibility of identifying diseases and their follow-ups. Pneumonia is one of the most widespread morbid and mortal diseases in the world that has radiographic similarities between bacterial and viral etiologies, making it difficult to diagnose.

Objective: This study aimed to compare the two deep learning models, ResNet-50, which is a convolutional neural network, and ViT-B/16, which is a transformer-based model in classifying images of chest X-rays based on bacterial pneumonia, viral pneumonia, and normal ones and also establishing the interpretability of their predictions.

Methods: A sdataset of X-rays of the chest containing bacterial, viral and normal X-rays was used. During preprocessing, normalization and augmentation was performed. ResNet-50 and ViT-B/16 models were trained, optimized and validated with stratified folds. Evaluation metrics included accuracy, F1-score, sensitivity, specificity, and AUC. Explainability was assessed using Grad-CAM to visualize clinically relevant regions.

Results: ResNet-50 achieved an accuracy of 87.2% with a macro F1-score of 0.86, while ViT-B/16 outperformed with an accuracy of 90.1% and a macro F1-score of 0.88. Bacterial pneumonia and normal classes were reliably detected, whereas viral pneumonia remained the most challenging category. Grad-CAM confirmed that both models focused on lung regions corresponding to pathological abnormalities, with ViT-B/16 demonstrating broader contextual attention.

Conclusion: Transformer-based deep learning architectures provide superior performance and interpretability compared to traditional CNNs, underscoring their potential to enhance disease detection and monitoring in medical imaging.