BossCE AI Blog

AI Image Review for Dentists and Doctors: What the Evidence Shows and How to Use It Safely

A Second Look at Imaging: BossCE AI article illustration
Purpose-built caries AI and general-purpose chatbots perform very differently on radiographs. What the studies show, how automation bias catches experienced readers, and a safe workflow for using AI as a second look.

By the BossCE Editorial Team.

Key Points

  • Purpose-built dental AI for proximal caries on bitewings performs well in pooled studies, but positive predictive value varies widely, so every positive finding needs clinical confirmation [1].
  • When clinicians read bitewings with AI support, sensitivity for proximal caries rises, especially for early lesions, while specificity is largely preserved [2].
  • General-purpose chatbots that accept images are a different tool. On dental radiographs they have missed most periapical lesions and obturation errors, and have described teeth that were not there [4][5].
  • Readers at every experience level follow wrong AI suggestions. Use AI as a second look after your own read, not before it [6][7].

Two Kinds of Image AI

“AI image review” covers two very different tools, and the evidence for one does not transfer to the other.

Purpose-built detection software is trained on labelled radiographs for a defined task, such as marking proximal caries on bitewings. It is tested against a reference standard for that task.

General-purpose multimodal models (chat assistants that accept an uploaded image) are trained for broad language and vision tasks. BossCE uses a general-purpose multimodal model (Google Gemini). It has not been trained or validated on a specific radiograph dataset, it does not segment anatomy, and it cannot reliably measure distances or lesion sizes on an uploaded JPEG.

Purpose-built detection AI General-purpose multimodal AI
Task One defined finding, e.g. proximal caries Open-ended description and discussion
Evidence base Diagnostic accuracy studies and meta-analyses [1][2] Small single-centre studies; results vary by model version [3][4][5]
Main strength Raises sensitivity for early lesions when used alongside a clinician [2] Structuring a review, listing differentials, explaining findings
Main weakness False positives can prompt unnecessary treatment [1] Low sensitivity for subtle findings and hallucinated features [4][5]

What the Evidence Shows: Purpose-Built Tools

A 2024 meta-analysis of 21 studies on approximal caries on bitewings reported pooled AI sensitivity of 0.94 and specificity of 0.91 [1]. The positive predictive value ranged from 0.15 to 0.87 and the negative predictive value from 0.79 to 1.00. The authors concluded that AI is highly reliable at excluding healthy surfaces, but that positive findings should be verified by a dentist to prevent unnecessary treatment [1].

A 2026 systematic review asked the more practical question: does a clinician do better with AI than without it? Seven studies covering 1,802 bitewings and 6,400 proximal surfaces were included, with dental students, general dentists and postgraduate clinicians as readers. AI assistance increased sensitivity, particularly for early lesions, and specificity fell by no more than 3.4%. Three studies reported better agreement between observers. Changes in reading time were modest and inconsistent [2].

What the Evidence Shows: General-Purpose Chatbots

General-purpose models are improving quickly, and results depend heavily on the model version and the task.

  • Radiology quiz cases. On 95 RSNA 2024 Case of the Day questions, which combine images and text, OpenAI o1 scored 59%, comparable to two senior radiologists (58% and 66%). GPT-4o scored 54%, Gemini 1.5 Pro 36% and Llama 3.2-90B-Vision 33% [3]. Quiz cases with a written stem are not the same as an unselected clinical image.
  • Periapical radiographs after root canal treatment. In 271 radiographs, ChatGPT-5 had specificity up to 99.3% for normal findings but sensitivity of 10.5%–28.6% for periapical lesions, 9.0%–22.7% for voids, 13.7% for short obturations and 0% for long obturations. Its overall accuracy of 54.0%–63.2% was significantly below that of general dentists (76.0%–85.6%) [4].
  • Third molars on panoramic radiographs. ChatGPT-4 and Microsoft Copilot scored 0.22 and 0.28 out of 1 for describing third molar position, and both hallucinated features, such as classifying absent teeth [5].

In the endodontic study the main failure was missed disease rather than over-calling [4]. A reassuring answer from a chatbot does not rule out disease.

The Human Factor: Automation Bias

In a prospective experiment, 27 radiologists read mammograms with a purported AI system that gave a deliberately wrong BI-RADS category on some cases. When the suggestion was wrong, the proportion of correct ratings fell from 79.7% to 19.8% for inexperienced readers, from 81.3% to 24.8% for moderately experienced readers and from 82.3% to 45.5% for very experienced readers [6]. Experience reduced the effect but did not remove it.

WHO’s 2024 guidance on large multi-modal models lists automation bias among the risks, alongside false, inaccurate, biased or incomplete statements, and sets out more than 40 recommendations for governments, technology companies and healthcare providers [7].

How to Use AI Image Review Safely

  1. Read the image yourself first. Record your findings before you look at any AI output, so the AI checks your read rather than steering it [6].
  2. Anonymise. Crop out names, IC or ID numbers, dates of birth, record numbers and faces before uploading.
  3. Give context. State the modality (bitewing, periapical, OPG, chest PA), the area of interest and the relevant history.
  4. Ask for structure, not a verdict. Ask it to work through a checklist and list differentials, rather than asking “what is wrong with this?”
  5. Ignore measurements. Disregard any millimetre values, bone-loss percentages or lesion sizes from a general-purpose model.
  6. Treat “normal” as unproven. Low sensitivity means a negative answer carries little weight [4].
  7. Confirm clinically. Use clinical examination, sensibility testing, comparison with previous images or further imaging. Where a formal radiology report is required, AI does not replace it.

Copy-Paste Prompts

Dental: bitewing

“This is an anonymised bitewing of the right posterior teeth in a 34-year-old with no symptoms. I have already noted [your findings]. Work through a systematic review: interproximal surfaces, restorations and margins, crestal bone and any visible periapical areas. Say which areas need a closer clinical look and which you cannot assess. Do not estimate measurements.”

Dental: a finding you have already spotted

“On this anonymised periapical of a lower first molar I can see a radiolucency at the mesial root apex. List the differentials, the clinical tests that would distinguish them, and when CBCT would be justified.”

Medical: chest radiograph

“This is an anonymised PA chest radiograph of an older adult with a chronic cough. Walk me through a systematic review: airways, lungs, pleura, heart, mediastinum, bones, soft tissues and review areas. List what should be confirmed with clinical findings, blood tests or a formal radiology report.”

Ask BossCE AI

Ask BossCE AI to structure your review of an anonymised image, or to explain a finding to a patient in plain language. For the underlying guidelines, browse the BossCE Library. Our article on periapical radiolucencies sets out a structured diagnostic approach. Never upload identifiable patient information; see our Privacy Policy.

Sources

  1. Carvalho BKG, Nolden EL, Wenning AS, et al. Diagnostic accuracy of artificial intelligence for approximal caries on bitewing radiographs: a systematic review and meta-analysis. J Dent. 2024;151:105388. doi:10.1016/j.jdent.2024.105388
  2. Salehizeinabadi M, Math SY, Ameli N, Demirturk H, Pacheco-Pereira C. Effectiveness of human-artificial intelligence collaboration for interproximal caries detection using bitewing radiographs: a systematic review. Oral Surg Oral Med Oral Pathol Oral Radiol. 2026. doi:10.1016/j.oooo.2026.07.005
  3. Hou B, Mukherjee P, Batheja V, Wang KC, Summers RM, Lu Z. One year on: assessing progress of multimodal large language model performance on RSNA 2024 Case of the Day questions. Radiology. 2025;316(2):e250617. doi:10.1148/radiol.250617
  4. Almutairi W, Alnasser M, Alharbi H, Austah O. Diagnostic accuracy of ChatGPT-5 in evaluating root canal treatment and periapical pathosis on periapical radiographs. Restor Dent Endod. 2026;51:e29. doi:10.5395/rde.2026.51.e29
  5. Silva TP, Andrade-Bortoletto MFS, Alencar-Palha C, et al. Evaluation of ChatGPT-4 and Microsoft Copilot for third-molar assessment on panoramic radiographs. J Imaging Inform Med. 2026;39:4686–4694. doi:10.1007/s10278-025-01805-y
  6. Dratsch T, Chen X, Rezazade Mehrizi M, et al. Automation bias in mammography: the impact of artificial intelligence BI-RADS suggestions on reader performance. Radiology. 2023;307(4):e222176. doi:10.1148/radiol.222176
  7. World Health Organization. Ethics and governance of artificial intelligence for health: guidance on large multi-modal models. 18 January 2024. who.int

This article is for qualified healthcare professionals and is not a substitute for clinical judgement or the full guideline. Check the current version of each source before applying it.

Share this articleWhatsAppLinkedInFacebook
Have a question about this topic?Ask BossCE AI for a sourced answer in seconds, with numbered references you can check.
Share the Post:

Related Posts

Free for doctors, dentists and students. Guests get 1 question a day. A free account gives you 20 a day, saved answers, Academy cases and a CPD log.

Create a Free Account