Thorax MapNet: Attention-Based Architecture with Anatomical Priors for Disease Classification

Abstract: The detection of thoracic diseases through chest medical images presents difficulties because of two main issues. The existing deep learning models include YOLOv8 and Faster R-CNN and U-Net-based CNN classifiers which achieve good results but these models struggle with two main issues. The study introduces A2-YOLOv8-ViT as a hybrid deep learning framework that integrates YOLOv8s real-time object detection capabilities with Vision Transformers global feature learning abilities. The proposed model uses spatial and channel attention mechanisms to enhance feature representation while highlighting important clinical areas and it uses a feature fusion strategy to combine local CNN features with transformer-based global dependencies for better detection results. The system uses an adaptive thresholding technique to solve class imbalance problems while boosting prediction accuracy. The proposed framework achieves 99% accuracy through its experimental results while showing better precision and recall and mean average precision results than traditional methods. The system produces interpretable outputs which include bounding boxes and confidence scores and severity indications that enable efficient and accurate clinical decision-making for thoracic disease diagnosis.

Keywords: Thoracic Disease Detection, A2-YOLOv8-ViT, Vision Transformer, Deep Learning, Medical Image Analysis, Attention Mechanism, Chest X-ray, Object Detection, CNN, Adaptive Thresholding.