Hybrid Framework for Multi-Disease Chest X-Ray Diagnosis Using Vision Transformers with Label Noise Correction and Uncertainty Calibration

The most accessible radiological technique for thoracic disease detection and diagnosis uses chest X-ray imaging as its primary method. The deployment of automated chest X ray analysis systems faces two main obstacles because of untrustworthy labels in large datasets and unpredictable predictive confidence levels. The research proposes a hybrid system which combines Vision Transformer (ViT) architecture with methods to handle noisy labels and produce accurate probability estimates for multiple disease diagnosis in chest X-ray images. The system trains on CheXpert and NIH ChestX-ray14 datasets while using Co-Teaching and DivideMix noise-handling methods and self-supervised pretraining to enhance feature resistance against supervision errors. The framework uses temperature scaling and Monte Carlo dropout as post-hoc methods to enhance confidence reliability without compromising discriminative performance. The system aims to reach performance levels that match or exceed traditional CNN and standard ViT models in AUROC and mAP and F1 score metrics. The system reduces the effects of untrustworthy labels while generating meaningful confidence scores which doctors can understand. The model produces Grad-CAM++ explanations to assist doctors in understanding its decision-making process. The hybrid system works to develop AI systems which deliver both exact results and safe operational readiness for real-world chest X-ray decision support systems.