Differential-fusion YOLO for MB2 canal detection and recognition in cone-beam computed tomographic images

Abstract

This study aimed to develop and validate an artificial intelligence (AI)-based system, called Differential-Fusion YOLO (DF-YOLO), for MB2 canal detection and recognition in cone-beam computed tomography (CBCT) images. DF-YOLO is a dual-branch convolutional architecture that integrates differential feature extraction with adaptive fusion to highlight subtle anatomical variations between adjacent slices. A clinical CBCT dataset comprising 148 microscopically confirmed MB2-positive cases (7,804 axial slices) was retrospectively collected and divided into training ( n = 116), validation ( n = 16), and test ( n = 16) groups. Although all included patients had clinically confirmed MB2 canals, only a limited number of slices from each case contained visible MB2 canal structures; the remaining slices were therefore classified as slice-level negative samples. Model performance was benchmarked against state-of-the-art detectors and further examined through ablation and human–AI comparison studies. For the human–AI comparison, 20 microscopically confirmed MB2-positive cases were included. In addition, an exploratory patient-level analysis was conducted using the 16 MB2-positive cases from the test set and 10 additional MB2-negative cases. DF-YOLO yielded an internal slice-level precision of 83.6%, recall of 87.4%, mAP@0.5 of 92.0%, and mAP@0.5:0.95 of 53.6%. In the exploratory patient-level analysis based on the positive-slice ratio, the model achieved an AUC of 0.881. Using a threshold of ≥ 0.20, the patient-level sensitivity was 93.8%, specificity 60.0%, accuracy 80.8%, positive predictive value (PPV) 78.9%, negative predictive value (NPV) 85.7%, and F1-score 85.7%. In the exploratory human–AI comparison, DF-YOLO achieved higher recall and faster inference than manual slice-level annotation. However, the limited number of positive cases in the clinician reader dataset precludes general conclusions about its diagnostic performance in routine clinical practice. The proposed DF-YOLO approach showed promising internal performance for MB2 canal detection on CBCT ROI slices. The exploratory patient-level analysis showed high sensitivity but only modest specificity in a small supplemental MB2-negative cohort. Larger, more rigorously curated negative-case datasets, external validation, and clinically representative cohorts are needed before clinical implementation can be considered.

The authors' abstract, as published at the source. BMC Oral Health, 2026 · DOI ↗