AIThatThinksLikeaDentist
DentalMind analyzes all 5 X-ray modalities in one pipeline, clusters findings per tooth, and gives your team ranked treatment prompts — not a list of bounding boxes.
Part of a broader research focus on trustworthy AI for healthcare — read the blog.
Dental AI promised less work. It delivered more review.
Three structural problems with today's dental imaging AI — and why they keep dentists from trusting the output.
Bounding boxes, not decisions
Current AI tools give you bounding boxes. You still have to think.
Single-modality blind spots
Bitewing only. OPG only. Never the full picture across modalities.
False positives everywhere
Dentists stop trusting the tool when every scan flags noise.
MadClip: four components, one coherent output
Each finding passes through all four before it ever reaches a dentist's screen.
Cross-slice 3D awareness
Sees neighbors, not just one slice
Consistency filter
Kills false positives before they reach you
Per-tooth clustering
Tooth 26: caries + bone loss = one compound finding
Treatment prompt
Ranked options, not just a label
One pipeline. All five modalities.
Every modality runs through the same shared encoder before branching into a specialized detection head.
Bitewing
YOLOv8x-seg
Caries accuracy: 87–95%Panoramic
YOLOv11 + HierarchicalDet
mAP: 0.848Periapical
Swin-T Hybrid
Accuracy range: 70–99%Full-Mouth Series
Shared DentVFM-2D
Per-series findings: Up to 18 imagesCBCT 3D
nnU-Net v2
Dice score: 0.87Datasets trained
Modalities supported
False positives filtered by C2
Compound finding patterns
How DentalMind stacks up
Against the incumbent dental imaging AI vendors.
| Feature | Pearl / Overjet / Videa | DentalMind |
|---|---|---|
| All 5 modalities | ||
| Per-tooth clustering | C3 | |
| Treatment prompt | C4 | |
| FP reduction | Clinical layer | Image-levelC2 |
| Open source backbone | MMKD-CLIP | |
| 2D + 3D joint train |