What Does Dental and Orthopedic Imaging Annotation Involve?

The Dental and Orthopedic AI Market: Why Annotation Demand Is Growing

The global dental AI market was valued at USD $3.8 billion in 2024 and is projected to reach USD $8.2 billion by 2030, growing at a CAGR of 13.7% (Grand View Research, 2024). The primary drivers are AI-assisted caries detection, periodontal assessment, implant planning, and OPG analysis — all of which require large volumes of annotated radiographic data. An estimated 3.5 billion people globally suffer from untreated oral disease (WHO Global Oral Health Status Report, 2022), creating significant demand for AI-based screening tools that can extend diagnostic capacity in regions with limited dental workforce.

The orthopedic imaging AI market is similarly expanding. The global AI in orthopaedics market was valued at USD $760 million in 2024 and is projected to reach USD $1.7 billion by 2029 (MarketsandMarkets, 2024), with fracture detection and joint replacement assessment AI as the largest application segments. Emergency department fracture triage AI — which assists radiographers and emergency physicians in prioritising radiograph review — is the fastest-growing application, driven by radiographer workforce shortages in Australian and UK public health systems.

Our dental and orthopedic imaging annotation service covers the full annotation stack for both domains — from basic radiograph labelling to specialist-reviewed diagnostic annotation for regulatory submissions.

Dental X-ray Annotation: Tooth Numbering, Caries Staging, and Pathology Localisation

Dental imaging annotation begins with tooth identification — assigning each visible tooth a code from the FDI World Dental Federation two-digit notation (ISO 3950). The first digit specifies the quadrant (1 = upper right, 2 = upper left, 3 = lower left, 4 = lower right) and the second specifies the tooth position (1–8 from central incisor to third molar), giving 32 permanent tooth identifiers from 11 to 48. Primary dentition uses quadrant codes 5–8, with codes from 51–85. Annotation protocols must specify how partially erupted teeth, impacted third molars, and supernumerary teeth are handled — these are among the most common sources of tooth-level annotation disagreement.

Caries detection and staging uses the International Caries Detection and Assessment System (ICDAS), which grades each tooth surface from 0 (sound) to 6 (extensive dentinal caries). The clinically critical thresholds for AI training data are:

Beyond caries, dental annotation tasks include periapical lesion delineation (polygon annotation of radiolucent lesions at root apices, indicating pulp necrosis or chronic infection), periodontal bone loss measurement (distance from the cementoenamel junction to the alveolar crest in millimetres, per tooth), root fracture detection, root canal morphology classification, and implant position and peri-implant bone level assessment.

OPG (orthopantomogram) panoramic radiograph annotation requires whole-image context — the annotator must assess the full dentition, mandibular condyles, maxillary sinuses, and hard palate simultaneously. OPG annotation guidelines must specify how image distortion artefacts (common in the anterior region of OPGs due to projection geometry) affect lesion sizing measurements.

Orthopedic Fracture Annotation: The AO/OTA Classification System

Orthopedic fracture annotation for AI training uses the AO/OTA Fracture and Dislocation Classification as the reference taxonomy. The system assigns each fracture a three-segment code: the first segment identifies the bone (1 = humerus, 2 = radius and ulna, 3 = femur, 4 = tibia and fibula, 5 = spine, 6 = pelvis and acetabulum, 7 = hand, 8 = foot); the second identifies the anatomical location (1 = proximal, 2 = diaphysis, 3 = distal); and the third identifies the fracture type (A = simple/two-part, B = wedge/three-part, C = complex/multifragmentary).

For emergency triage AI — the most common deployment context for orthopedic fracture detection — annotation guidelines simplify the AO/OTA system to what is visible on standard emergency radiographs (two views: AP and lateral). Each annotation includes: fracture present/absent (binary), anatomical region (per a defined list for the body area), laterality (left/right), fracture type (A/B/C), and image quality flag. Occult fractures — cases where cortical disruption is subtle and visible only on one projection, or requires a dedicated oblique view — must have explicit handling rules in the annotation protocol, as these cases are the single largest driver of false negatives in fracture detection AI.

Paediatric fractures require Salter-Harris classification for growth plate injuries (Type I through V, from physeal widening to crush injury), which requires specific annotator training because growth plate anatomy differs markedly from adult cortical bone. Stress fractures, insufficiency fractures in osteoporotic bone, and pathological fractures through metastatic lesions all require separate annotation sub-categories with specific detection rules.

Bone Density, Joint Replacement, and Soft-Tissue Annotation

Beyond fracture detection, orthopedic imaging annotation covers three additional clinical AI applications. Bone density assessment annotation supports osteoporosis screening AI — measuring cortical thickness (the width of the cortical shell at standardised measurement points, most commonly the second metacarpal and the femoral neck), trabecular texture pattern classification, and vertebral fracture severity grading using the Genant Semi-Quantitative (SQ) method (Grade 0 = normal, 1 = mild, 2 = moderate, 3 = severe). Osteoporosis affects an estimated 200 million women globally (International Osteoporosis Foundation, 2023) and drives substantial demand for opportunistic osteoporosis AI on routine radiographs.

Total joint replacement assessment annotation supports AI systems that evaluate component position, alignment, and wear in postoperative radiographs. Annotation tasks include acetabular cup inclination angle measurement, femoral stem alignment, polyethylene liner wear measurement, cement mantle quality grading, and periprosthetic bone remodelling pattern classification. These measurements, performed manually by radiologists and orthopaedic surgeons on serial postoperative radiographs, are a strong candidate for AI automation — but require precise annotation of geometric measurements rather than simple classification labels.

For X-ray annotation and MRI annotation supporting musculoskeletal soft-tissue applications — rotator cuff tear grading, meniscal tear classification, ligament injury assessment — annotation tasks shift from radiograph-based measurement to segmentation and classification tasks on multi-sequence MRI. These tasks require orthopaedic radiologist or specialist annotation due to the complexity of soft-tissue pathology characterisation on MRI.

Case Study: Fracture Detection AI for an Australian Emergency Department Network

A digital health company was developing a musculoskeletal fracture triage AI for deployment across a network of 14 Australian emergency departments. The intended use was second-reader AI: radiographs taken in the ED are processed by the AI immediately after acquisition, and cases where the AI detects a probable fracture are flagged for priority radiologist review. The clinical goal was to reduce time-to-diagnosis for fractures by surfacing high-probability cases ahead of the radiologist queue.

The company had trained an initial model on 18,000 musculoskeletal radiographs annotated by a general annotation vendor using non-specialist annotators. Evaluation against a 2,200-image radiologist-graded test set showed overall fracture detection sensitivity of 68.4% at 85% specificity — well below the performance threshold required for clinical deployment. The type C (complex/comminuted) fracture false negative rate was 52.7%, and occult fractures (cortical disruption not visible on the primary AP view) were missed in 61.3% of cases.

Root cause analysis: Diagnostic review of 400 false negative cases from the initial model revealed two primary annotation failure modes. First, the non-specialist annotators had no protocol for occult fractures: cases where the AP view showed only subtle cortical blurring were annotated as fracture-absent (negative), even though the lateral view — which was included in the image set — showed clear cortical disruption. Second, type C comminuted fractures were systematically under-annotated because the annotation guideline required a single bounding box per fracture, and annotators were placing boxes over the most obvious fragment rather than covering the full fracture extent including satellite fragments.

Our approach: We developed a dual-view annotation protocol in collaboration with two Australian orthopaedic radiographers and one orthopaedic surgeon. The protocol required both views to be reviewed simultaneously before any annotation decision, with explicit rules for the five most common occult fracture patterns at each anatomical site (scaphoid waist, radial head, fifth metatarsal base, lateral process of talus, and tibial plateau). Comminuted fractures received a polygon annotation covering the entire fracture zone including satellite fragments, with a separate minimum-bounding-box measurement recorded for model training on the detection head. All 22,000 radiograph pairs were annotated by trained radiographers, with 100% orthopaedic surgeon review on AO type B and C cases and 20% random sampling on type A cases.

Annotator Qualification: Who Should Annotate Dental and Orthopedic Images?

Annotator qualification for dental and orthopedic imaging AI follows a tiered model based on task complexity. At the base tier, trained radiography annotators — individuals who have completed dental radiograph or musculoskeletal radiograph annotation training, including a structured protocol examination and a calibration set — can reliably perform tooth identification on high-quality bitewing and periapical radiographs, binary fracture detection on standard AP/lateral pairs, and image quality grading.

The specialist tier is required for tasks where clinical judgement is the primary determinant of label accuracy. For dental AI: ICDAS caries staging beyond Grade 2 (where the enamel-dentin boundary ambiguity requires dentist-level training), periapical lesion differential diagnosis, and root canal morphology classification for endodontic AI. For orthopedic AI: AO/OTA type B and C fracture classification, growth plate Salter-Harris staging in paediatric cases, pathological fracture identification through metastatic lesion, and joint replacement component assessment.

Inter-annotator agreement benchmarks vary by task. Dental caries detection (binary present/absent, bitewing radiographs) targets Cohen's kappa above 0.80 between trained annotators. ICDAS caries grade agreement between dentists on the same radiographs is typically kappa 0.65–0.75 in the literature, reflecting genuine clinical ambiguity at the enamel/dentin boundary. Fracture detection (binary) targets kappa above 0.85 between trained radiographers for major fractures, and above 0.70 for occult fractures (where clinical ambiguity is higher). See our guide to Cohen's kappa in annotation quality for how to interpret these thresholds.

Regulatory Compliance: HIPAA, FDA 21 CFR Part 11, and TGA SaMD

Dental and orthopedic AI intended for clinical use requires annotation datasets that comply with medical device regulations in each target market. In Australia, the TGA regulates diagnostic dental and orthopedic AI as SaMD under the Therapeutic Goods Act. Class IIa devices (decision support with a clinician in the interpretation path) require a technical file including annotator qualifications, inter-annotator agreement statistics, and a description of the annotation protocol used to establish the reference standard.

Patient data transferred to annotation teams must be de-identified to Australian Privacy Act standards (equivalent to HIPAA Safe Harbor for US-origin data). De-identification for dental and orthopedic radiographs includes removing patient name, date of birth, and Medicare/insurance identifiers from DICOM headers, as well as any embedded identifying text (patient name overlaid on the radiograph image). Our guide to FDA 21 CFR Part 11 annotation documentation covers the full provenance logging requirements for regulatory submissions.

For radiology annotation across modalities, annotation audit trails must record: annotator identifier, annotation timestamp, protocol version applied, QA reviewer identifier and review timestamp, and the disposition of any cases flagged as ambiguous. For dental and orthopedic AI specifically, the audit trail must also record the image quality assessment result — because model performance specifications are typically quoted on images above a defined quality threshold, and the training data annotation must document which images fell below threshold and were excluded from the training dataset.

Frequently Asked Questions