Radiograph AI for caries: more lesions found, and more drilling?
In a randomized trial, AI support helped dentists find more enamel caries on bitewings. The same AI support also led to more invasive treatment decisions. We put that trial next to two FDA clearances and the vendors' own launch numbers, and kept the three apart.
Written Oct 31, 2026 (ET) PubMed abstract, two FDA 510(k) records, two vendor pages, Dental Tribune, a roundtable summary and ADA News fetched Oct 31, 2026
Partly. In one randomized trial AI support raised sensitivity for early enamel lesions without a significant change in specificity, but it also raised invasive treatment decisions. Vendor launch percentages are their own claims.
- 1 Vendor case study
- 2 Operator-reported
- 3 Organization evaluation
- 4 Randomized trial
What operators / vendors claim
Reported by the named group or vendor. Treated as claims until independently verified.
- Pearl (Second Opinion). Pearl's March 8, 2022 press release says the product passed "four separate clinical studies, each featuring a radiographic dataset of over 2,000 images and a reader group of 86 expert dentists and dental radiologists". It also says readers using the AI "accurately identified 36% more lesions" than readers without it. Claim · vendor
- Overjet (Caries Assist). Overjet's May 16, 2022 post says it analyzed "more than 7,000 tooth surfaces" and that, with the aid of Caries Assist, "dentists were able to detect 32% more tooth surfaces containing caries". It also says the tool "reduces the rate of missed lesions (false negatives)". Claim · vendor
- An operator's view. A summary of Becker’s Oct 8, 2026 roundtable recap says Affinity Dental Management, described as an early Overjet adopter, "saw more treatment planning once doctors used AI diagnostics". It gives no percentage. Claim · operator (secondary)
- "First" and "#1" language. Both companies describe themselves with firsts and category leadership in their own announcements. We treat that wording as marketing, not as a finding. Claim · vendor positioning
The 36% and 32% figures are not comparable with each other. Each comes from a different company's study design and reference standard, and neither press release gives specificity or false-positive rates.
What we independently checked
What Signal Desk could confirm from public sources this pass — not a clinical study.
- A randomized trial exists, and its main result is modest. Mertens et al. (Journal of Dentistry, Dec 2021) ran a cluster-randomized crossover trial: 22 dentists each read 20 bitewings, 10 with and 10 without a commercial AI tool (dentalXrai Pro). With AI, mean AUC was 0.89 vs 0.85 without, and sensitivity was 0.81 vs 0.72. Specificity was not significantly affected. Checked · PubMed abstract
- The gain was in enamel lesions only. The abstract says the sensitivity increase "was found for enamel, but not early or advanced dentin lesions". Checked · PubMed abstract
- More invasive treatment decisions. Higher sensitivity "came with an increase in non-invasive, but also invasive treatment decisions". The authors conclude AI "may also increase invasive treatment decisions". Checked · PubMed abstract
- Both clearances are real FDA records. K210365 (Second Opinion, Pearl) was decided Mar 4, 2022 and K212519 (Overjet Caries Assist) on May 10, 2022. Both are "Substantially Equivalent" decisions under regulation 892.2070, product code MYN. Checked · FDA 510(k) database
- No shared validation set yet. ADA News (Feb 3, 2025) reports that ADA Technical Report 1109:2025 calls for an independent dataset held by a third party "that is not an AI manufacturer" so that algorithms can be compared. Checked · ADA News
What did not go well / limits
Gaps, caveats, and anything we could not verify. Numbers without a fetched source stay here.
- The trial tested a different product. The randomized trial used dentalXrai Pro, not Pearl or Overjet. Its results do not transfer automatically to other tools. Limit · different product
- Small, image-only study. 22 dentists and 140 bitewings, read outside a real patient visit. No clinical exam, patient history, or long-term outcomes were included. Limit · setting
- Possible developer link. Dental Tribune (Nov 3, 2021) reports that dentalXrai Pro was co-founded at Charité by the senior author and later spun off as a start-up. We did not read the paper's conflict-of-interest statement, so we cannot say how it was handled. Unchecked · COI statement
- Vendor percentages not re-checked. We did not open the FDA decision summaries' performance sections or any vendor study protocols. The 36% and 32% figures stay as claims. Unchecked · vendor study data
- Clearance is not a ranking. A 510(k) "substantially equivalent" decision is not a head-to-head comparison. It does not show better patient outcomes or fewer unnecessary fillings. Limit · regulatory meaning
Owner takeaway: ask any vendor for sensitivity and specificity on an outside dataset. Also decide in advance how your team will treat AI-flagged enamel lesions (monitor, remineralize, or restore) so that higher detection does not quietly turn into more drilling.
Sources
- Artificial intelligence for caries detection: Randomized trial (Mertens, Krois, Garcia Cantu, Arsiwala, Schwendicke)
- Study highlights how artificial intelligence can be used for detection of caries
- Past the Hype: Where AI Is Delivering Real ROI for Dental Groups
- 510(k) Premarket Notification K210365: Second Opinion (Pearl, Inc.)
- 510(k) Premarket Notification K212519: Overjet Caries Assist (Overjet, Inc.)
- FDA Clears World’s First AI Software to Read Dental X-Rays
- Overjet Receives Second FDA Clearance, Adding Overjet Caries Assist to the Industry’s #1 Dental AI Platform
- What are the standards for AI use in dentistry?
← All case studies · Related directory category · How we label evidence · Editorial case for practice owners evaluating AI tools. Not a vendor advertisement.