Canonical text (the exact bytes fingerprinted)

TRUVACE RECORD VERSION
record: TRV-2026-2005
version: 1
kind: certified
reason: Certified into the record
timestamp: 2026-10-20T10:42:37.260971Z
status: published
lens: trace
sector: health
headline: Differential-fusion YOLO for MB2 canal detection and recognition in cone-beam computed tomographic images
dek: 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 re…
gain_title: The DF-YOLO model detected MB2 canal structures on CBCT slices with 87.4% recall and, in an exploratory patient-level test, 93.8% sensitivity, and it ran faster than manual slice annotation.
problem_title: In the exploratory patient-level analysis, specificity was only 60.0% in a small MB2-negative cohort, so the model may flag canals in cases without them.
trace_subject: AI detection of MB2 canals in CBCT images
gain_reading: The DF-YOLO model detected MB2 canal structures on CBCT slices with 87.4% recall and, in an exploratory patient-level test, 93.8% sensitivity, and it ran faster than manual slice annotation.
gain_evidence: DF-YOLO achieved higher recall and faster inference than manual slice-level annotation | the patient-level sensitivity was 93.8%
problem_reading: In the exploratory patient-level analysis, specificity was only 60.0% in a small MB2-negative cohort, so the model may flag canals in cases without them.
problem_evidence: high sensitivity but only modest specificity in a small supplemental MB2-negative cohort
quick_read: Researchers built DF-YOLO, a dual-branch convolutional network that compares adjacent CBCT slices to pick out subtle anatomical differences. They trained and tested it on retrospectively collected scans from patients with microscopically confirmed MB2 canals.

The results suggest automated MB2 detection could support clinicians, but the evidence is early and internal. The test sets were small, negative cases were limited, and no external validation has been done, so routine clinical performance is unknown.
limitation: The study is retrospective with small test and reader datasets and few negative cases. It lacks external validation, so the authors say it is not ready for clinical implementation.
tag: Dual reading
key_points: The dataset had 148 microscopically confirmed MB2-positive cases covering 7,804 axial slices. | Cases were split into 116 training, 16 validation, and 16 test cases. | Internal slice-level results were precision 83.6%, recall 87.4%, mAP@0.5 92.0%, and mAP@0.5:0.95 53.6%. | The exploratory patient-level analysis gave an AUC of 0.881 and used a positive-slice ratio threshold of 0.20 or higher. | The authors say larger negative-case datasets, external validation, and representative cohorts are needed first.
rundown: DF-YOLO is designed to highlight small differences between neighboring slices through differential feature extraction and adaptive fusion. It was benchmarked against state-of-the-art detectors and examined through ablation and human-AI comparison studies.

Because every included patient had a confirmed MB2 canal, most slices lacked a visible canal and were treated as slice-level negatives. The patient-level analysis added 10 MB2-negative cases to the 16 positive test cases.

At the patient-level threshold, accuracy was 80.8%, PPV 78.9%, NPV 85.7%, and F1-score 85.7%. The human-AI comparison used 20 positive cases, and the authors say this small number precludes general conclusions about routine diagnostic performance.
sources:
- peer_reviewed | BMC Oral Health | https://doi.org/10.1186/s12903-026-10156-1 | 2026-10-19
prev: 0000000000000000000000000000000000000000000000000000000000000000

sha256

13ae68a053fcca0c4b926bd2ab2c837c36af1b0864403c70359dfaa8af336161

previous

0000000000000000000000000000000000000000000000000000000000000000