Hybrid Vision Transformer and CNN Framework for Multi-Disease Pulmonary Diagnosis
R. Sriramkumar, K. Selvakumar, J. Jegan
Abstract
Chest X-rays often display modest patterns which makes interpreting the results both a clinical and technical challenge when it comes to diagnosing lung disorders. This paper proposes a multi-label classification for five principal thoracic conditions using a deep learning model that integrates the hierarchical vision transformer Swin with ConvNeXt, a modern convolutional neural network. The model is trained and evaluated on the BIMCV-COVID19+ dataset incorporating sophisticated preprocessing techniques such as lung segmentation, CLAHE enhancement, and data augmentation to improve model resilience. Various experiments demonstrate that our hybrid model surpasses performance achieved by state-of-the-art CNN and transformer-based baselines with an overall classification accuracy of 93.4% and AUC of 0.964. Moreover, interpretable tools like Integrated Gradients and Grad-CAM++ confirm with supporting evidence the clinical suitability of the model's predictions enhancing windows into what factors most influenced its outputs bolstering trust in its use within medical imaging disciplines. In conclusion, these findings assertation that radiology workflows would benefit from an integrated consideration of attention and convolutional techniques for supporting automated decisions enables powerful means for diagnosis automation regarding pulmonary disease detection.