Deep learning-based detection of carotid artery atheromas in panoramic radiographs

Abstract

Stroke is among the leading global causes of mortality and long-term disability resulting from neurological sequelae, with carotid atherosclerosis being one of its main etiological mechanisms. This condition often progresses silently, which limits the use of standard diagnostic examinations such as Doppler ultrasound and computed tomography angiography (CTA) for population screening. However, calcifications located at the carotid bifurcation may be incidentally visualized on panoramic dental radiographs, making this exam a potential opportunistic screening tool due to its wide availability, low cost, and coverage of the C3–C4 region, where atheromas usually manifest. In this context, the present study aimed to develop and evaluate a deep-learning-based method using the MobileNetV2 architecture for the automatic detection of carotid atheromas in panoramic radiographs, aiming to optimize the clinical value of this routine exam as a screening tool. A total of 378 publicly available and fully anonymized Regions of Interest (ROIs) were used, cropped in the carotid region (640×320 px) and divided into training, validation, and test sets (264/57/57). The images underwent normalization, grayscale channel replication, and real-time data augmentation, and class imbalance was addressed using weighted loss. The model was implemented using MobileNetV2 with pretrained weights and a two-stage training scheme, consisting of initial backbone freezing followed by partial fine-tuning (~70%), with batch normalization and dropout (0.3). On the independent test set, the model achieved robust and well-balanced performance, with 94.7% accuracy, 95.7% sensitivity, 94.1% specificity, and AUC and AUPRC values of 0.963 and 0.968, respectively, using a threshold optimized by Youden’s J index. These results demonstrate strong discriminative capability and reinforce the potential of panoramic radiography combined with deep learning as an opportunistic screening tool for carotid atheromas. The proposed method shows promising performance and supports the use of routine panoramic radiographs for opportunistic screening, enabling the early referral of patients for specialized vascular evaluation and strengthening the interface between oral and systemic health.