The review followed the Preferred Reporting Items for Systematic Reviews and Meta-Analyses 2020 statement.( 12)

Studies were eligible when they were original investigations that developed or evaluated an artificial intelligence, machine learning, or deep learning model for detecting, staging, segmenting, or measuring periodontal or alveolar bone loss on panoramic or periapical radiographs and reported at least one quantitative performance measure against a reference standard. Accepted measures were sensitivity, specificity, area under the receiver operating characteristic curve, Dice or intersection over union, accuracy, F1-score, intraclass correlation coefficient, or mean measurement error. Studies whose primary input was cone-beam computed tomography or an intraoral photograph were excluded, as were reviews, editorials, and studies reporting no quantitative metric. The publication window was 2018 to 2025. The task definitions were deliberately broad. Detection comprised studies that identified the presence of bone loss, staging comprised studies that assigned a severity class, measurement comprised studies that estimated bone level or the percentage of bone loss, and segmentation comprised studies that delineated the bone or landmark pixels, with studies combining tasks assigned to all relevant categories.

PubMed, OpenAlex, and Crossref were searched for records published between 2018 and 2025. Queries combined terms for artificial intelligence, periodontal or alveolar bone loss, and dental radiography, with several synonymous strings run in each source to maximize recall. The PubMed query set comprised three strings: ("artificial intelligence" OR "deep learning" OR "machine learning") AND ("periodontal bone loss" OR "alveolar bone loss" OR periodontitis) AND (radiograph OR panoramic OR periapical) AND (detection OR staging OR diagnosis OR classification); "deep learning" AND ("bone loss" OR periodontitis) AND ("panoramic radiograph" OR periapical); and "convolutional neural network" AND (periodontal OR "alveolar bone") AND (radiograph OR radiography). Equivalent free-text queries were adapted to the OpenAlex and Crossref syntax, and each source returned up to 50 records per query. Records were deduplicated by digital object identifier and by title before screening.

Screening proceeded in two stages. Title and abstract screening excluded clearly irrelevant records, and uncertain records were carried forward rather than discarded. The title and abstract screen applied the eligibility criteria conservatively, so that any record whose relevance could not be determined was retained for full-text assessment. Full text was available for a subset of reports; for the remainder, eligibility and data extraction relied on the abstract.

For each included study the following items were extracted: first author and year, country, model architecture, imaging projection, sample size, reference standard, task, and reported diagnostic performance. Values were recorded as reported, without recalculation, and when a study reported separate values for different lesion severities or imaging projections, the stratum-specific values were retained. When a study reported staging accuracy for the 2017 classification, the stage-specific values were recorded, and when a study reported agreement statistics, the intraclass correlation coefficient or the kappa was retained alongside the accuracy. No values were imputed, and studies for which the performance could not be extracted were excluded at the eligibility stage rather than included with missing data.

Risk of bias and applicability were assessed with QUADAS-2 across patient selection, the index test, the reference standard, and flow and timing.( 13) Because most studies reported no prospective registration and no independent external validation, risk of bias was summarized narratively by domain.

A meta-analysis was not performed because the studies reported heterogeneous tasks, metrics, and reference standards. Detection studies reported sensitivity and specificity, staging studies reported classification accuracy, and measurement studies reported agreement coefficients and error in millimeters, and these cannot be combined into a single estimate without strong and unjustified assumptions. Findings were synthesized narratively, grouped by imaging projection and by task, and heterogeneity was described rather than statistically tested.

The searches retrieved 350 records, of which 113 duplicates were removed and 237 unique records were screened (Figure 1). Title and abstract screening excluded 187 records, leaving 50 reports assessed for eligibility. Of these, 12 were excluded: 7 because no extractable quantitative metric was reported, 2 because the imaging input was not a two-dimensional dental radiograph, 2 because the outcome was not periodontal bone loss, and 1 because the report was not an original study. Thirty-eight studies were included (38).

The included studies spanned 2018 to 2025, with the largest number published in 2025 (11 studies) and 2024 (10 studies), and the earliest report in 2018 (Table 1). Panoramic radiographs were the most common input, used in 23 of 38 studies, and periapical radiographs in the remaining 15. In terms of task, 20 studies performed detection of bone loss, 9 performed severity staging, 7 performed quantitative measurement of bone level, and 2 performed pixel-level segmentation. The detection studies used bounding boxes or classification labels as their output, and the segmentation and measurement studies used pixel masks or landmark coordinates, so the format of the model output, and therefore the metric used to evaluate it, differed across the studies.

Convolutional neural networks dominated the architectures. Object detectors based on the YOLO family and region-based detectors such as Faster R-CNN and Mask R-CNN were the most frequent.( 14-20) Segmentation networks based on U-Net were also common,( 21-23) and several studies applied classification networks such as VGG-16, Inception, and ResNet.( 24-27) A minority used transformer architectures(28,29) or a multimodal large language model.(30)

The reference standard was expert annotation in most studies, whether a single periodontist, a calibrated panel, or a consensus of specialists. Clinical periodontal assessment served as the reference in a subset,( 22,31,32) and a few studies measured agreement against manual annotation.(33,34) Sample sizes ranged from fewer than 100 images(30,35) to more than 20,000 radiographs,(28,29) and only a minority reported external validation on an independent cohort.( 18,36)

The studies also differed in the anatomical landmarks they measured. Most models localize the cementoenamel junction and the alveolar bone crest and express bone loss as the ratio between the two, which mirrors the clinical definition of radiographic bone loss.(21,34,37) A few studies measured bone loss at specific sites, such as approximal surfaces,(33) or combined the radiographic measurement with a clinical stage assigned according to the 2017 classification.(31,32) The choice of landmark and of unit of analysis, whether site, tooth, or patient, differs across studies and complicates the comparison of their reported errors.

A further source of variation is the landmark on which the measurement is anchored. The cementoenamel junction is difficult to localize on panoramic radiographs, and some studies anchor the measurement on the tooth apex or on adjacent structures instead, which changes the resulting bone level percentage. The studies that reported the lowest agreement with the reference standard cited this difficulty explicitly,(33,38) and the studies that reported the highest agreement used a controlled periapical projection in which the landmarks are more visible.( 22,34)

The studies also originated predominantly from Asia and Europe, with China, South Korea, and Turkey contributing several reports each, and a smaller number from North America, South America, and the Middle East.( 3,22,25,26,32,39) The geographic concentration is relevant because panoramic machines, exposure protocols, and patient demographics differ across regions, and a model validated in one setting may not transfer to another. Sample sizes also grew over time: early studies used a few hundred radiographs,(16,38) whereas recent studies used several thousand, and two large periapical studies shared a dataset of more than 21,000 radiographs.( 28,29)

Across QUADAS-2 domains, the reference standard was the principal source of concern. When bone loss is labelled by expert readers, the model is evaluated against human judgement rather than against the true bone level, and inter-reader agreement for bone loss is only moderate, so the target itself contains error. A minority of studies used clinical assessment or a calibrated consensus as a stronger reference,(31,32) and none used a criterion independent of human readers. Patient and image selection raised further concerns because cohorts were assembled retrospectively from single institutions in most studies, and few reports described consecutive or random sampling. Flow and timing were generally acceptable, but blinding was reported inconsistently, and only a minority held out an independent test set.( 18,32,36)

The applicability of the findings to routine practice is also uncertain. Most models were trained on images from a single institution and a single acquisition device, and performance measured internally may not transfer to a different panoramic machine or to a different population. The two studies that reported an external validation cohort illustrate this. One found the external performance lower than the internal estimate,( 18) and the other reported specificity and accuracy that were retained on an independent set.( 36) These mixed results suggest that external validation, where performed, tends to reduce the apparent accuracy, which is the expected direction for any model evaluated outside its training distribution. Taken together, the risk-of-bias appraisal indicates that the reported accuracy should be interpreted as an upper bound. The combination of expert reference standards, retrospective convenience sampling, and internal test sets inflates apparent performance relative to what would be observed in routine practice, and only a handful of studies took the steps, such as external validation or a clinical reference, that would temper this inflation.

Reported performance varied widely across tasks and studies (Table 2). Among detection studies, sensitivity ranged from 0% for severe bone loss in one study that found the model unable to detect advanced lesions( 39) to 100% for total alveolar bone loss in a U-Net study.( 21) Thirteen studies reported sensitivity and ten reported specificity. Specificity ranged from 66.2%( 29) to 98%,( 19) and most detection studies that reported both measures obtained sensitivity and specificity at or above 80%.(2,16,26,27,40) Eight studies reported the area under the curve, which is a threshold-independent measure and therefore better suited to comparing models that used different operating points.

The area under the receiver operating characteristic curve was reported in eight studies and ranged from 0.70 for a classification study of intrabony defects(27) to 0.951 for a U-Net detection study.(21) Intermediate values were reported for detection on periapical radiographs, with area under the curve between 0.884 and 0.918 in a large multi-model study,(28,29) and for detection on panoramic radiographs, with a value of 0.91.(17) One multi-center study reported an area under the curve of 0.942 for the model, which exceeded the 0.856 achieved by the periodontal specialists who served as the reference.(32)

Among the detection studies, the reference standard conditioned the reported accuracy. Studies that used a clinical or consensus reference reported sensitivity and specificity above 0.85,(31,40,41) whereas studies that used a single expert reader or a manual annotation reported more moderate agreement.(2,35) One study that directly compared the model with the clinicians who defined the reference reported that the model exceeded the clinicians on the area under the curve, a result that is possible only when the reference is imperfect.(32) This pattern suggests that the ceiling of reported accuracy is set by the reference standard, and that comparisons between studies with different references should be interpreted with caution.

Staging and measurement studies reported lower and more variable accuracy than detection studies. Staging accuracy ranged from 0.58 for a severity classification in periapical radiographs(38) to 0.968 in a study that combined deep features with a random forest classifier,(42) with intermediate values near 0.77 to 0.89,(4-6) and a deep CNN reported accuracy of 0.728 overall but 0.97 for stage III lesions.(37) Measurement studies reported intraclass correlation coefficients from 0.601(33) to above 0.94,(22) and mean bone-level errors around 5% to 11% of the bone length.(6,38) Segmentation studies reported Dice coefficients from 0.91(23) and intersection over union up to 0.849,(23) with one study reporting a Dice of 0.91 and staging area under the curve near 0.90.(34)

The staging studies illustrate the spread more concretely. A Mask R-CNN pipeline reported accuracy of 0.896 in one center and 0.824 in another,(5) an ensemble combining YOLOv8, Mask R-CNN, and TransUNet reported accuracy of 89.45% with an intraclass correlation of 0.806 against three periodontists,(6) and a U-Net plus YOLOv4 pipeline reported accuracy of 0.77.(4) The variability among these staging studies, all of which target the same clinical classification, is larger than the variability among detection studies, which is consistent with the greater difficulty of the task.

The segmentation studies, which are a prerequisite for quantitative measurement, reported pixel-level agreement that was generally high. A U-Net enhanced with a super-resolution network reported a Dice coefficient of 0.91 and intersection over union of 0.849,( 23) and a YOLOv8 model reported sensitivity of 90% to 100% and specificity of 96% to 98% for tooth, cementoenamel junction, and bone level segmentation.( 19) These segmentation results support the feasibility of automatic landmark localization, even though the downstream measurement accuracy remains more variable.

The segmentation studies are also informative in their own right. They show that networks can reliably separate tooth, bone, and the cementoenamel junction, which are the anatomical components required for any quantitative analysis. The highest agreement was reported for the bone and the cementoenamel junction, with sensitivity and specificity at or above 96%, and the lowest for the tooth itself, which is the least ambiguous structure to a human observer.( 19) This pattern, in which the network is most precise on the structures that are hardest for humans to localize consistently, is what makes the technology attractive despite its current limits.

The strongest detection results came from models applied to panoramic radiographs with large training sets. A U-Net trained on 1,121 panoramic radiographs reported sensitivity of 1.0 and an area under the curve of 0.951,(21) and a commercial system evaluated against a three-clinician consensus reported accuracy of 0.98 for binary and 0.99 for multiclass detection.(41) A DeNTNet model trained on 12,179 panoramic radiographs reported an F1-score of 0.75, which exceeded the 0.69 achieved by the clinicians who annotated the data.(3) These results indicate that the ceiling for detection, at least on curated internal test sets, is close to the ceiling of the reference standard itself.

The contrast between detection and staging is clearest within studies that performed both. A pipeline that segmented bone, tooth, and the cementoenamel junction reported a Dice coefficient of 0.91 for segmentation but a case diagnosis accuracy of 0.85,( 34) and another reported high segmentation agreement but staging accuracy of 0.77.(4) The pattern across these studies is consistent: the model localizes the relevant anatomy well, but the final severity assignment, which depends on a ratio and a threshold, is where accuracy is lost.

A commercial software evaluated against a consensus of three clinicians reached accuracy of 0.98 for binary and 0.99 for multiclass bone loss detection,( 41) whereas a multimodal large language model evaluated on panoramic images was judged insufficient for clinical use on the basis of its bone level estimates.( 30) The wide spread between these endpoints reflects differences in the model, the reference standard, and the difficulty of the task rather than a single underlying capability.

Detection performance also differed between panoramic and periapical radiographs. On panoramic radiographs, a U-Net model reported sensitivity of 1.0 and an area under the curve of 0.951 for total alveolar bone loss,(21) a Faster R-CNN model reported sensitivity of 0.84 and specificity of 0.88,(16) and a transfer learning model based on GoogleNet Inception reported sensitivity of 0.94 and specificity of 0.88.(26) On periapical radiographs, an ensemble of a detector, a classifier, and a segmentation network reported accuracy of 0.97 for bone level detection,(15) and a pipeline combining MobileNet-v2 with YOLOv8 reported accuracy of 0.88 with sensitivity of 0.7 to 1.0.(43) An early deep learning study of periodontally compromised teeth on periapical radiographs reported accuracy of 81.0% for premolars and 76.7% for molars,(44) and a two-stage CNN reported accuracy of 0.762 for radiographic bone loss classification.(45)

The measurement studies reported agreement against manual or clinical annotation, and here the results diverged. A segmentation network that delineated the bone, the tooth, and the cementoenamel junction reported a Dice coefficient of 0.91 and a staging area under the curve between 0.89 and 0.90,(34) and a cascading SegFormer model reported intersection over union of 0.789, an F1-score of 0.867, and a correlation with clinician measurements above 0.85.(46) A U-Net plus YOLOv9 pipeline reported accuracy of 95.62% and intraclass correlation above 0.94 against clinical assessment.(22) By contrast, an hourglass network reported a mean bone level error of 10.69% and a severity staging accuracy of only 0.58,(38) and a CNN evaluated against manual annotation reached an intraclass correlation of 0.601.(33) Pixel-level agreement is therefore high, whereas agreement with the clinical ground truth is more variable. The measurement studies that reported the highest agreement used a clinical or consensus reference and a controlled projection, whereas those that used a single manual annotation reported lower agreement, which again points to the reference standard as a determinant of the measured accuracy.

The field also evolved over the review period. The earliest studies used comparatively small datasets and reported moderate accuracy.( 2,3,44) Studies published in 2024 and 2025 applied more recent architectures to larger datasets and reported higher accuracy, including several above 90%.( 21,22,41) The reporting of measurement error and of external validation also increased in the more recent reports,(18,36) although external validation remains uncommon.