A Hybrid YOLOv8-ResNet50 Architecture for Enhanced Cardiomegaly Prediction from Chest X-rays

Abstract

Abstract.

Objective: This study aims to develop a reliable deep learning architecture for predicting cardiomegaly by integrating the ResNet-50 backbone into the YOLOv8 object detection framework, overcoming the challenges of detecting subtle anatomical variations and low-contrast features often found in chest radiographs.

Methods: This study used a publicly available chest X-ray dataset, with rigorous data annotation to establish ground truth for the heart and thoracic cavity regions. Preprocessing included resizing input images to 640×640 pixels, automatic orientation correction, and an 80:20 data split between training and testing. Real-time data augmentation was applied to the training set. The ResNet-YOLOv8 hybrid model was trained for 150 epochs with optimized hyperparameters (learning rate, momentum, weight decay, loss weight), and performance was evaluated using metrics such as mAP, precision, recall, and confusion matrix results.

Results: The experimental results show that the proposed architecture achieves high accuracy in detecting cardiomegaly, with mAP50-95 of 0.7578, precision of 0.9955, recall of 0.9962, F1 score of 0.9959, and inference latency of only 4.5 ms/img. This model is more optimal than the standard YOLOv8 variant in both accuracy and computational efficiency.

Innovation: The integration of ResNet-50 into YOLOv8 significantly improves feature extraction capabilities for chest X-ray images, enabling the recognition of fine anatomical details with high precision. This innovative hybrid approach advances automated cardiomegaly detection, offering potential for large-scale, real-time implementation in clinical settings and contributing to the development of advanced AI-powered diagnostic tools.