A BMC-Net model for the recognition and segmentation of mandibular canal bifurcation

To accurately identify the anatomical structure and trajectory of the mandibular canal, helping dentists avoid surgical risks and develop effective treatment plans, this study aimed to develop a deep learning model for rapidly diagnosing and identifying mandibular canal bifurcation.

We collected a total of 160 reported case images from the PubMed and Web of Science databases. Of these, 140 images were allocated to a training set and 20 to a testing set, to develop the BMC-Net deep learning model. Performance was evaluated using the dice similarity coefficient (DSC), area under the curve (AUC), intersection over union (IoU), recall, precision, and a confusion matrix, with comparisons made against the UNet model. To determine clinical utility, a comparative analysis with clinicians was performed, focusing on AUC, sensitivity, specificity, and time efficiency.

The BMC-Net model achieved a DSC of 0.9704, AUC of 0.9458, IoU of 0.9412, recall of 0.9458, and precision of 0.9829 in the training set, with significant improvements in the testing set, where it reached a DSC of 0.9877, AUC of 0.9965, IoU of 0.9539, recall of 0.9569, and precision of 0.9976. Compared to clinicians, the BMC-Net model achieved an AUC of 0.9636, sensitivity of 0.9314, and specificity of 0.9461, and detected mandibular canal divergence in just 0.1004 s, significantly faster than the average clinician's recognition time of 95.5 to 164.75 s (P < 0.05).

This groundbreaking model's utility markedly improves the accuracy of clinical diagnoses on mandibular nerve canal bifurcations.