AI-Powered Medical Imaging: Enhancing Diagnostic Accuracy in Radiology

Abstract

Medical imaging plays a pivotal role in disease diagnosis, particularly in identifying thoracic conditions such as COVID-19 and pneumonia using chest X-rays. However, the complexity and variability in medical images often limit the effectiveness of traditional convolutional approaches, which struggle with capturing long-range spatial dependencies. To address these limitations, we propose a hybrid deep learning architecture that integrates the strengths of Vision Transformers (ViT) for global spatial feature extraction with recurrent neural network variants—RNN, GRU, and LSTM— for temporal sequence modeling. In our framework, the ViT-Base-Patch16-224 model, pretrained on ImageNet-21k, is employed to extract high-dimensional embeddings from 16×16 image patches. These embeddings are then passed through RNN-based architectures to model sequential dependencies among the patches, enabling a richer and more context-aware representation. The proposed models were evaluated on a balanced dataset of 6939 posteroanterior chest X-ray images, comprising three classes: COVID-19, Pneumonia, and Normal. Extensive experiments revealed that the ViT-LSTM model achieved the highest classification accuracy of 96.33%, outperforming both ViT-GRU (95.51%) and ViT-RNN (93%). These results highlight the effectiveness of combining global attention mechanisms with sequential learning for robust and interpretable medical image classification. Our findings suggest that ViT-LSTM is a promising approach for enhancing diagnostic precision in radiology and could support clinical decision-making in real-world healthcare settings.

Ähnliche Arbeiten