Detection of Parathyroid Adenoma in Primary Hyperparathyroidism using Deep Learning Models

The precise identification of parathyroid adenoma in CT imaging is essential for prompt medical management, although it continues to be difficult due to complex anatomical changes. This work assesses the diagnostic efficacy of four fine-tuned deep learning architectures, Xception, VGG19, ResNet50, and Inception V3, trained and validated on 4DCT images for the classification of parathyroid adenomas. The dataset used for the training of these models was privately obtained and contained 24 positive verified cases and 24 negative verified cases of parathyroid adenoma by the radiologists. Among the models, Xception demonstrated superior performance, achieving an AUC of 0.88, accuracy of 77.3%, and a balanced precision-recall trade-off (F1 = 0.7655) at a threshold of 0.45, with high sensitivity (recall = 0.74) crucial for minimizing missed diagnoses. The study underscores the impact of architectural design on model performance, with Xception’s depthwise separable convolutions proving particularly effective for spatial hierarchy learning in CT imaging. By establishing a framework for deep learning in medical imaging, this work advances diagnostic precision in parathyroid adenoma detection through optimized architecture selection and threshold tuning, offering valuable insights for clinical application.