Mucosight AI

Mucosight AI is a Japanese cloud-based artificial intelligence program that analyzes photographs of the inside of the mouth and provides findings to help dentists decide whether to recommend consultation at a higher-level medical institution. PMDA's approved-software register lists Morita Tokyo Manufacturing as the approval applicant, while the system grew from University of Osaka research supported by Nvidia. Japan's Ministry of Health, Labour and Welfare approved it on February 24, 2026 as a Class II medical-device program under approval number 30800BZX00053000.^{[1][2][3][10]}

The approved purpose is referral-decision support in primary dental care. It is not an autonomous screening service for patients, a pathology system, or a substitute for a dentist's or physician's diagnosis. The University of Osaka says the system presents findings with features associated with oral cancer, leukoplakia, benign tumors, and stomatitis, but final diagnosis remains the responsibility of a clinician.^[1] As of August 17, 2026, the product had regulatory approval but had not yet been documented as commercially launched. The university and Nvidia described a domestic service release as planned for around late August.^[1][5]

Regulatory status and intended use

The product's Japanese sales name is 口腔粘膜画像解析システム Mucosight AI. Its general device name is 病変検出用口腔内画像診断支援プログラム, which describes a lesion-detection intraoral image diagnostic support program. A February 16, 2026 MHLW notice created that general name as a Class II category, and PMDA's JMDN database assigns it code 71148002 and GHTF rule 10.^[2][3]

PMDA defines the category as software that further processes intraoral images obtained from general-purpose equipment for diagnostic use and can detect lesion candidates.^[2] The product-specific approval followed on February 24. The University of Osaka identifies the approval as ministerial approval following PMDA review, gives approval number 30800BZX00053000, and states the intended use as supporting whether a primary dental provider should recommend consultation at a higher-level institution. PMDA's approved-software register independently lists the same sales name, general name, approval number, February 24 date, and Morita Tokyo Manufacturing as the applicant.^[10] A trade report on MHLW's February approval register also pairs the product and date with Morita Tokyo Manufacturing.^[4]

This distinction limits what the approval establishes. Mucosight AI supplies image findings for referral consideration. It does not establish a definitive diagnosis of cancer, determine histology, or replace examination and testing by an oral specialist. The university also warns that evidence has not been established for diseases or anatomical sites outside the approved range.^[1]

Development and data

University of Osaka says Shin-Ichiro Hiraoka's group began the underlying work in 2017. The university led the clinical research, review of images and labels, evaluation of model output, and management of intellectual property. Nvidia contributed technical support and computing for model development. Morita Tokyo Manufacturing handled productization, regulatory work, and the regulated manufacture-and-market role, according to the university.^[1] PMDA's register lists the company as the approval applicant, but the reviewed public records do not separately identify a physical manufacturing site.^[10]

The project involved oral-surgery groups at Tohoku University, Ohu University, the National Center for Global Health and Medicine, Fujita Health University, Asahi University, Kumamoto University, and other institutions. Specialists reviewed lesion location, disease classification, image quality, and consistency with clinical information before images were prepared for model development.^[1] A Tohoku University disclosure for its part of the project describes a 2019 to March 2023 research period, de-identified transfer of data to Osaka, and the collection of photograph-diagnosis pairs plus pathology and other clinical information. That disclosure describes one site's contribution to the broader research, not the final composition of the approved product's training data.^[6]

Nvidia reported in August 2026 that the multicenter effort collected approximately 40,000 intraoral images. It said specialists selected and quality-checked images, mainly those with pathologically confirmed diagnoses, for training and validation.^[5] The public account does not give the exact number of patients or images used, class balance, demographic composition, camera distribution, rules for multiple images from one patient, or the division of institutions and patients among training, validation, and test sets. The 40,000 figure therefore describes collection scale, not an independently characterized test cohort.

The University of Osaka says the work received support from an AMED translational research program and a university Innovation Bridge grant. It also discloses that the university owns related patents and has a patent-license agreement with Morita Tokyo Manufacturing.^[1]

Workflow and output

Mucosight AI accepts a digital-camera image of oral tissue and analyzes it on a cloud server.^[1] Its described finding classes are oral cancer, leukoplakia, benign tumor, and stomatitis. The university lists the tongue, gingiva, floor of mouth, buccal mucosa, and palate as examples of oral sites covered within the approved range.^[1]

According to Nvidia's vendor account, a dentist uploads an image through a web browser. Processing takes about 30 seconds. If a disease similarity measure exceeds a threshold, the interface draws a bounding box around the candidate area; the box color indicates one of the four finding classes. Nvidia also says the result can be attached to a referral letter.^[5] The reported 30-second time is a vendor description, not an independent service-level measurement.

The interface output should not be read as a calibrated cancer probability unless future product documentation establishes that interpretation. The public description is of localization and class-coded similarity, followed by human consideration of whether referral is appropriate. A specialist may still need direct examination, further imaging, biopsy, and pathology to determine the diagnosis.^[1][5]

Evidence and limitations

Publicly accessible sources do not identify the architecture or version of the model in the approved product. Earlier project materials describe deep learning, Nvidia GPUs, and development tools, but do not establish a particular neural-network architecture, parameter count, production runtime, calibration method, or model-update policy for Mucosight AI.^[7] The system can be described as image-analysis AI within computer vision, but a more specific architecture such as a particular convolutional network or object detector has not been documented.

An Nvidia release from 2020 reported sensitivity and specificity of at least 95 percent for malignant-tumor and stomatitis detection in an earlier four-class research model.^[7] That announcement predates the approved device by more than five years and did not provide sufficient cohort, split, confidence-interval, or external-validation details. The figures cannot be treated as the performance of the 2026 Mucosight AI product.

The 2026 university and vendor releases provide no product-specific sensitivity, specificity, area under the curve, predictive values, false-positive or false-negative rates, confidence intervals, reader-study results, or subgroup analyses.^[1][5] No accessible product-specific peer-reviewed validation paper, package insert, or public PMDA review report was found by the August 17 cutoff. Public sources likewise do not report a prospective trial showing that use of Mucosight AI improves referral accuracy, time to diagnosis, cancer stage, treatment outcomes, mortality, quality of life, or clinician workload. Regulatory approval should not be presented as proof of any of those benefits.

The broader oral-imaging literature supports caution about transferring results between datasets. A 2026 clinician-centered review led by Hiraoka found that prospective designs and clinical-impact endpoints remained scarce in oral-oncology AI. It identified differences in cameras, lighting, blur, mucosal glare, pigmentation, dental restorations, lesion mimics, and clinical settings as potential sources of domain shift. The review called for external validation, clinician-in-the-loop evaluation, calibration, and measurement of patient-relevant outcomes before routine adoption.^[8] An independent 2023 systematic review similarly found heterogeneous study methods and limited generalizability, and noted that oral diagnosis extends beyond image appearance alone.^[9] Neither review evaluates Mucosight AI itself.

The University of Osaka says the next phase will use real clinical experience to refine appropriate photography, referral criteria, specialist assessment, and patient explanation. Nvidia's account also describes evaluation of patient usefulness as a future step.^[1][5] These plans indicate ongoing evidence generation rather than a completed demonstration of clinical benefit.

Development data were reportedly kept on University of Osaka workstations, with Nvidia staff accessing them remotely.^[5] That statement concerns research data handling. Because the approved product performs cloud analysis, deployment raises separate questions about hosting, encryption, retention, deletion, access logging, incident response, and secondary use of submitted images. Accessible sources did not document those controls. Public pricing, reimbursement, distributor identity, and commercial terms were also undisclosed before the planned launch.

References

1. ^University of Osaka Graduate School of Dentistry, *Oral mucosal disease AI research moves toward social implementation*, originally announced August 4, 2026 and updated August 17, 2026, dent.osaka-u.ac.jp/page-19045
2. ^Pharmaceuticals and Medical Devices Agency, *JMDN entry: Lesion-detection intraoral image diagnostic support program*, accessed August 17, 2026, std.pmda.go.jp/...stdDB_jmdn_resr.cgi
3. ^Ministry of Health, Labour and Welfare, *Notice Iyakuhatsu 0216 No. 3*, February 16, 2026, mhlw.go.jp/...T260216I0050.pdf
4. ^MTJ One, *February approval register includes six new in vitro diagnostics*, March 9, 2026, mtj.jiho.jp/...2618
5. ^Nvidia, *Oral cancer early-detection support with AI: Oral mucosal image-analysis system developed from NVIDIA-supported research moves toward implementation*, August 4, 2026, prtimes.jp/...000000645.000012662
6. ^Tohoku University Graduate School of Dentistry, *Public disclosure for research protocol 2019-3-15: Research to establish an AI method for diagnosing oral mucosal disease*, accessed August 17, 2026, dent.tohoku.ac.jp/...2019-15.pdf
7. ^Nvidia, *University of Osaka Graduate School of Dentistry and NVIDIA begin joint deep-learning research for oral-cancer early detection*, November 12, 2020, nvidia.com/...ve-accuracy-ai-oral-cancer-detection
8. ^Hiraoka S-I, Kawamura K, Akiyama R, et al., *Artificial intelligence for diagnosis and triage in oral cancer: a clinician-centered narrative review*, International Journal of Clinical Oncology 31, 794-803, 2026, ir.library.osaka-u.ac.jp/...IntJClinOncol_31_794.pdf
9. ^Gomes RFT, Schuch LF, Martins MD, et al., *Use of Deep Neural Networks in the Detection and Automated Classification of Lesions Using Clinical Images in Ophthalmology, Dermatology, and Oral Medicine: A Systematic Review*, Journal of Digital Imaging 36, 1060-1070, 2023, pmc.ncbi.nlm.nih.gov/...PMC10287602
10. ^Pharmaceuticals and Medical Devices Agency, *List of software medical devices granted manufacturing and marketing approval*, as of March 31, 2026, pmda.go.jp/...000281616.xlsx