Effectiveness of lightweight convolutional neural networks for detecting the relationship between the mandibular third molar and the inferior alveolar canal on panoramic radiographs

Objective: This study aimed to develop and evaluate lightweight convolutional neural networks (CNNs) capable of automatically localizing the mandibular third molar (M3) and classifying its relationship with the inferior alveolar canal (IAC) on panoramic radiographs.Methods: A total of 609 panoramic radiographs (containing 899 M3s) were analyzed in two stages. First, 82 panoramic images (134 M3s) were used to fine-tune a pre-trained EfficientDet model for automatic M3 localization. The detected regions were standardized to include the IAC and preprocessed through resizing, contrast enhancement, and mirroring. For ground-truth labeling, the presence or absence of M3–IAC contact was determined by an experienced oral and maxillofacial radiologist based on established panoramic radiographic criteria. Second, a custom lightweight CNN was trained on 527 panoramic radiographs (765 M3s) to classify M3–IAC contact (contact = 1, no contact = 0). Model performance was compared with a pre-trained ResNet50 architecture using accuracy, sensitivity, specificity, precision, and F1 score.Results: The detection model achieved 100% accuracy with an intersection-over-union (IoU) of 87.9%. Compared to the ResNet50 benchmark model, the lightweight CNN demonstrated comparable overall accuracy (87.5%). However, the lightweight CNN outperformed ResNet50 in specificity (90.4% versus 86.9%) and precision (93.4% versus 88.7%), while ResNet50 exhibited a slightly higher mean sensitivity (88.3% versus 86.2%).Conclusions: Lightweight CNNs can achieve diagnostic performance comparable to large pre-trained networks while requiring less training time and computational power. The proposed model enables automated, efficient, and clinically feasible detection of the M3–IAC relationship on panoramic radiographs.