Transformer and generative models in dental and maxillofacial imaging: Narrative review of clinical evidence and future directions

Abstract

Purpose

This narrative review examined whether primary dental and maxillofacial imaging studies support, qualify, or counter the conclusions of previous systematic reviews of transformer and generative models, with findings assessed by topic rather than by review.

Materials and Methods

Structured PubMed searches covering 2019–2026 retained 159 records, and supplementary searches addressed topic-specific gaps. Dental and maxillofacial evidence was distinguished from medical transfer evidence. Thirty-six comparisons were then assessed for external validation, patient separation, comparator fairness, and clinical endpoints.

Results

In task-level architecture comparisons, previous reviews reported frequent first-place rankings for transformer models but limited external validation. Primary studies qualified that pattern: convolutional neural networks won the largest matched segmentation comparison, Swin-Unet rankings reversed across datasets, convolutional models led two direct cone-beam computed tomography comparisons, and one transformer retained high performance in external surface segmentation. Of 36 comparisons, 4 used independent external validation, 15 had a fair same-data comparator, and 15 included a diagnostic, reader, or clinical endpoint. Generative reviews reported image-quality gains, but heterogeneity exceeded 99% in a cone-beam computed tomography meta-analysis, and most dental metal-artifact studies had a high risk of bias. Crack visibility improved ex vivo; low-dose cone-beam computed tomography improved structure visibility without significantly changing treatment or referral decisions.

Conclusion

Primary evidence supports complementarity between convolutional neural networks and transformers and suggests early clinical translation, but it does not support architecture-wide transformer superiority. The directly observed gaps were matched comparisons, independent external testing, use-specific endpoints, and measurement of anatomy removed or introduced by pixel-modifying models.

Introduction

Dental and maxillofacial care uses panoramic and intraoral radiographs, cone-beam computed tomography (CBCT), cephalometric images, and intraoral optical scans to address different clinical questions. Across these modalities, artificial intelligence (AI) has been investigated for segmentation, detection, classification, landmark localization, registration, and image restoration.1,2,3,4

Recent systematic reviews of dental architecture comparisons reported that a vision-transformer-based model ranked first in 58% of 21 comparisons, that complementarity between convolutional neural networks (CNNs) and transformers was proposed, and that 18 of the 21 studies lacked external validation.5 Another review found that hybrid models consistently outperformed pure transformers, with a 5%–8% decrease in performance from internal to external validation.6 This narrative review asks whether source-checked primary studies support, restrict, or counter these review-level conclusions when task, dataset, comparator, and endpoint are considered. Rather than comparing the reviews one by one, it organizes the evidence by topic: architecture ranking, external validation, generative image quality, and clinical translation.

In the 112-study review, 91 architectures combined CNN and transformer components, whereas 23 were pure transformers.6 The primary studies analyzed here also included conventional CNNs, attention-CNNs, and geometric or graph-based methods. Architecture labels were retained as reported in the source papers to avoid imposing classifications not used by the original authors.

For the generative portion of the review, pixel-modifying applications were classified as super-resolution, denoising, harmonization, synthesis, CBCT-to-CT translation, or metal-artifact reduction. Image-quality results were analyzed separately from endpoints involving structure, readers, diagnosis, or treatment decisions.

This synthesis does not infer a general architectural hierarchy from scores obtained on different datasets. Instead, it identifies same-data comparisons, independent external tests, reader or decision endpoints, and direct tests of structure preservation, and then uses only those observations to derive unresolved research questions.

The main synthesis covers panoramic and intraoral radiography, CBCT and maxillofacial CT, cephalometry, and intraoral optical surface scans. Clinical photography provides only modality-specific oral evidence. Histopathology, cytology, fluorescence-lifetime imaging, tongue-image datasets, and unrelated medical applications are outside the core scope; medical or radiotherapy evidence is used only to clarify methods relevant to dental image formation.

Materials and Methods

Search strategy

PubMed was searched through the National Center for Biotechnology Information E-utilities interface, using esearch and efetch, in five rounds on July 6–7, 2026. Searches were limited to 2019–2026, English-language records, and human studies where applicable. The core transformer-specific title/abstract block combined transformer[tiab], “vision transformer”[tiab], “Swin transformer”[tiab], or self-attention[tiab] with dental, tooth/teeth, oral, maxillofacial, CBCT, or panoramic terms. Eight query families added specific transformer and generative architectures, imaging tasks, medical-transfer methods, and recent reviews. From round 2 onward, language-model, vision-language, and report-generation terms were excluded. Complete field-tagged queries, per-round counts, and query–to–PubMed identifier (PMID) logs are provided in the Supplementary Material.

Eligibility and study selection

After deduplication by PMID, 1,363 hits yielded 946 unique records; nine records with unusable metadata were removed. One reviewer screened the remaining 937 titles and abstracts. Records advanced when they described an identifiable in-scope imaging task and were original transformer or generative studies, task-matched CNN comparators, or directly relevant reviews or medical-transfer sources. Eligible tasks included segmentation, detection, classification, landmarking, registration, synthesis, translation, denoising, artifact reduction, and super-resolution. Transformer evidence required token self-attention; an attention gate or registration-only spatial transformer was not classified as a vision transformer. Records were excluded if they involved language or report generation without an imaging model, lacked an identifiable imaging task, addressed out-of-scope anatomy without direct dental-transfer value, or were too peripheral to use. CNN comparators were purposefully selected rather than obtained through a separate systematic CNN search; architecture comparisons are therefore exploratory.

At report assessment, relevance was scored from 0 to 5 as follows: 5, core dental evidence directly addressing the review question; 4, other directly relevant dental evidence; 3, relevant reviews or medical-transfer evidence; 2, peripheral but usable contextual evidence; 1, tangential evidence; and 0, off-topic or non-imaging evidence. Reports scoring 2 or higher were retained. Of 168 reports assessed, 167 were reviewed in full, and one panoramic-harmonization report was available only as an abstract; nine reports scored below 2, leaving 159 retained reports (Fig. 1). The abstract-only report did not support any numerical conclusion. No second screener or interrater agreement assessment was used.

Fig. 1. Study selection and composition of the evidence base.

Recent syntheses revealed recall gaps in the primary queries. Supplementary searches on August 4–5, 2026, therefore targeted image formation, segmentation and detection, and landmarking, measurement, and registration. Each query family used retmax = 1000 rather than the 6–30-record limits used in the primary search; families with more than 1,000 matches were truncated at that ceiling, so the supplementary searches were recall-oriented rather than exhaustive. Cached records were screened against the claims targeted by each round using titles and abstracts, and added numerical claims were limited to values stated in those abstracts. These results refined the topical narrative but were not added to the 159-record corpus, Figure 1 counts, or 36-study source-checked set; their strategies, ceiling flags, and results are reported separately in the Supplementary Material.

Clinical scope and provenance

Each record was assigned to one mutually exclusive provenance class. Dental-native studies examined the dentition, periodontium, tooth-bearing jaws, a dental treatment or appliance, or a dental population. Maxillofacial studies examined adjacent craniofacial targets without a specifically dental question. Transfer studies examined general medical imaging, radiotherapy, or phantoms and were not treated as evidence of dental performance.

The 159 retained records comprised 87 dental-native records, one maxillofacial record, and 71 transfer records; 81 were original studies, and 78 were reviews, perspectives, or dataset reports. Among the original studies, 52 developed a transformer or generative method, and 29 mainly served as CNN or attention-CNN comparators. Transfer evidence was retained only when it addressed a directly relevant method or safety requirement.

Optical and laboratory records remained in the audit trail but were kept outside the core synthesis. Findings from clinical photographs were not generalized to radiographic modalities.

Extraction and synthesis

Full texts were summarized with AI-assisted structured extraction using a fixed one-paper-per-record schema. Extracted fields included modality, target, architecture, sample size, split strategy, metrics, comparator, clinical endpoint, and a supporting quotation. The authors determined eligibility, provenance, design classification, interpretation, and wording, and they checked citations against the reference list. The Supplementary Material catalogs the search, extraction, classification, and verification records.

The synthesis was narrative because tasks, datasets, comparators, and metrics were too heterogeneous to pool. No formal risk-of-bias or certainty-grading instrument was applied. Comparative studies were instead assessed for cohort denominator and unit, patient-level separation, leakage, external validation, comparator fairness, metric definition, and diagnostic, reader, or clinical endpoints. Conclusions from prior systematic reviews were grouped by topic: task-level architecture ranking, external validation and transport, generative image quality, and clinical translation. These conclusions were then compared with source-checked primary studies. A conclusion was described as supported when the primary result and validation pattern agreed, qualified when task-matched evidence restricted its scope, countered when a matched comparison ranked another architecture first, and adjacent when the closest identified primary evidence concerned related but nonidentical material, modality, or task and therefore could neither directly support nor counter the conclusion. Scores from different datasets were not vote-counted, and an unobserved endpoint was treated as a gap in this corpus rather than as proof of absence from the entire literature.

Thirty-six comparative studies were checked against primary-source records: 22 dental-native studies and 14 declared transfer studies. Thirty-two were checked in full text, and four were assessed at the abstract level. Four used independent external validation, 15 had a fair same-data comparator, and 15 reached a diagnostic, reader, or clinical endpoint (Fig. 2A); one of the 15 fair same-data comparators was adjudicated from its abstract only. These counts describe this source-checked set, not all dental AI studies.

Fig. 2. Evidence characteristics in the source-checked sets. A. Appraisal of the 36 source-checked comparative studies. B. Pixel-modifying evidence from the primary search and from supplementary reviews.

Results

Convolutional, transformer, and hybrid architectures

U-Net is a convolutional encoder–decoder architecture with skip connections between matched resolution levels.7 Vision transformers divide an image into tokens and apply self-attention across those tokens,8 whereas Swin transformers calculate attention within shifted local windows.9 These definitions describe model structure but do not, by themselves, predict performance on a dental imaging task.

The reviewed hybrid models combined convolutional and transformer components. Foundation models provide promptable segmentation,10 and foundation-model priors contributed to top-performing entries in a public tooth-segmentation challenge.11 Individual adaptations were tested for tooth-related tasks in photographs and oral CBCT, but not in independent clinical-site tests.12,13

Terminology should be interpreted carefully. An attention-gated U-Net may still be a CNN, and a spatial transformer used for image warping is not a vision transformer. More importantly, the architecture label alone does not show whether patients were separated or whether testing used an independent scanner or clinical site.

Generative and restoration methods

Generative adversarial networks train a generator against a discriminator,14 whereas diffusion models learn to reconstruct images from added noise.15 Across the included dental and declared transfer studies, pixel-modifying methods were classified as super-resolution, harmonization, synthesis, CBCT–to–computed tomography (CT) translation, denoising, or metal-artifact reduction.

A systematic review of biomedical super-resolution reported that high-resolution outputs can introduce or exaggerate features absent from the input.16 The dental studies below were therefore separated according to whether they measured only image similarity or also evaluated structure, reader performance, diagnosis, or treatment decisions.

Segmentation

Segmentation reviews reported pooled Dice scores of approximately 0.93 for teeth and 0.694 for the mandibular canal; only 6.67% of studies in one tooth-segmentation review used external validation.17,18 A later U-Net meta-analysis estimated a pooled mandibular-canal Dice score of 0.84, with a 95% confidence interval of 0.68–0.92 and high heterogeneity.19

In panoramic radiography, Swin-Unet outperformed CNN baselines on a 100-image in-house set but underperformed those baselines on a public dataset, showing dataset-dependent rank reversal within the same study.20 The largest matched segmentation comparison trained two CNNs, two hybrid models, and two transformers under identical five-fold cross-validation on 1,881 panoramic and 4,314 bitewing images. CNNs performed significantly better for teeth, tooth structures, and caries; mean F1 scores were 0.89 versus 0.83 for teeth and 0.49 versus 0.32 for caries.21

In a separate six-model condyle benchmark, a transformer detector achieved the highest Dice score, 0.866, although all models fell within a narrow range of 0.819–0.866.22 A public challenge using 2,380 panoramic images and 330 CBCT scans reported gains from foundation-model priors and coarse-to-fine refinement compared with an nnU-Net trained only on labeled data.11 Neither result reverses the matched findings for tooth and caries segmentation because the targets and test sets differed.

For CBCT segmentation, a frequency-attention U-Net outperformed TransUNet and UNETR for mandibular-canal segmentation in 91 subjects; the authors attributed part of the transformer result to overfitting.23 In an external multi-architecture comparison, six models were trained on 105 pediatric scans and tested on 46 scans from another center. A convolutional nnU-Net variant ranked first externally (Dice, 0.846), whereas the transformer was less consistent for primary teeth.24 Thus, both direct CBCT comparisons in this synthesis ranked a convolutional or attention-convolutional model above the tested transformers.

For intraoral optical scans, a global-local attention model achieved a mean intersection over union (mIoU) of 92.14% internally and 91.73% on an external multicenter, multiscanner set.25 A 44-study meta-analysis pooled performance at 90.2% and identified the gingival boundary as the main residual error region.26 This finding provides direct external evidence for a transformer on surface scans, but no included study tested whether that result transfers to CBCT.

Detection and classification

A Faster R-CNN with a Swin backbone reached an area under the curve (AUC) of 0.977 for fracture detection in 6,404 panoramic radiographs. However, because the study did not train a matched CNN backbone, the contribution of Swin cannot be isolated.27 A 23-study facial-fracture review found that both convolutional and Swin-based models commonly exceeded 0.90 and that studies were predominantly retrospective, single-center designs without external or prospective validation.28

An attention-enhanced YOLOv8 model for second-molar lesions decreased from an internal F1 score of 0.926 to 0.781 on an independent external test.29 This result estimates external degradation for a CNN system, not a CNN–transformer comparison.

A hybrid Swin model reported an AUC of 0.922 for childhood caries on panoramic images in an internal test.30 A convolutional classifier for epithelial dysplasia in oral photographs was tested independently against clinicians and had an external AUC of 0.828.31 Because the second study used photographs and a different target, the two values cannot be used to rank the architectures.

A periapical-radiolucency meta-analysis identified 30 studies but could pool only six, and it graded certainty as low.32 In caries detection, Mask R-CNN reached a pooled sensitivity of 85.6% and an AUC of 0.95, compared with 71.7% and 0.84 for Faster R-CNN.33 This within-family difference indicates that the CNN label alone did not determine performance in that review.

Landmark localization, measurement, and registration

For cephalometric landmarking, one meta-analysis reported that 79.9% of landmarks were within 2 mm and rated the risk of bias as high in most studies;34 another reported that 79% were within 2 mm and 90% were within 3 mm, with I² = 99% and very low certainty.35 Both reviews noted repeated use of the International Symposium on Biomedical Imaging (ISBI) 2015 benchmark and variation in landmark definitions.34,35

For three-dimensional landmarking, a 37-study scoping review found that 56.8% of included studies met a 2-mm mean radial error threshold, 75.7% used fewer than 200 cases, and no accepted three-dimensional accuracy standard exists.36 One externally validated study used a lightweight 3D U-Net trained on 720 CT and CBCT cases and tested on 470 separate cases. Mean radial error remained below 1.3 mm overall and below 1.4 mm in difficult cases, with no significant internal–external difference.37

For transformer-based approaches, direct-coordinate vision transformers reduced mean radial error by more than 2 mm relative to earlier methods,38 and a cross-attention regression network reported a mean radial error of 0.98 mm and an 89.82% success rate at 2 mm on the ISBI test set.39 Both were single-center benchmark results and therefore establish competitiveness rather than preference.

For forensic age estimation on 15,195 panoramic radiographs, a hybrid model achieved a mean absolute error of 2.61 years, compared with 3.08 years for a CNN and 7.48 years for a pure transformer.40 This same-data result supports hybridization for that task but does not establish a general architecture rule.

For image registration, a meta-analysis of 22 CBCT-to-intraoral-scan registration studies identified 14 geometric methods, seven AI methods, and only one direct comparison.41 Automated registration for implant planning reached a median surface deviation of 0.04 mm in 51 seconds, compared with 840 seconds for manual registration.42 Graph-based palatal superimposition kept tooth-position deviation below 0.32 mm.43 No included dental registration study directly compared a self-attention model with these geometric, convolutional, or graph-based methods.

Summary of the task-level comparisons

The three task sections above report results separately for segmentation, detection and classification, and landmark localization, measurement, and registration. Table 1 consolidates these findings before the generative evidence is presented, with each statement restricted to the design and endpoint reported by the cited studies.

Table 1. Task-specific findings and limits demonstrated by the included evidence.

CBCT: cone-beam computed tomography, CNN: convolutional neural network

Dental-native generative evidence

The primary corpus contained four original dental-native pixel-output studies, all involving panoramic radiographs (Fig. 2B). Three were available in full text, and one was available only as an abstract.

Two super-resolution studies compared convolutional, adversarial, and transformer candidates on 888 and 1,714 images. The best reference-based models reached peak signal-to-noise ratios near 40 dB, but neither study measured lesion detection, diagnostic accuracy, or treatment decisions after restoration.44,45 Cross-scanner CycleGAN harmonization was reported only in an abstract and supplied no numerical claim used here.46 A diffusion model trained on 5,383 panoramic images produced images that 14 clinicians could not reliably distinguish from real images, whereas classifiers separated real from generated images almost perfectly; the study did not use an anatomical reference or test diagnostic improvement.47

Supplementary searches substantially changed the estimated size of the field. A 43-study dental and maxillofacial generative review found that 24 studies (55.8%) addressed image quality, 14 (32.6%) addressed image simulation, three (7.0%) addressed three-dimensional reconstruction, and two (4.6%) addressed interpretation or reporting. Objective metrics were used in 95.3% of studies, subjective assessment in 58.1%, and code or model accessibility had a high risk of bias in 81.4%.48 The four panoramic studies in the primary corpus therefore describe primary-search retrieval rather than the size of the generative field.

A review of 10 dental CBCT metal-artifact studies reported improved structural similarity, peak signal-to-noise, and error metrics and favored transformer-based or physics-informed dual-domain models over image-only U-Nets. However, most included studies had a high risk of bias, and diagnostic impact had not been validated.49 Among 18 dental GAN studies, 72% had a low risk of bias in all four QUADAS-2 domains, but only three (17%) had low applicability concern.50 A 37-study CBCT enhancement meta-analysis included 27 CBCT-to-CT studies and estimated gains of 3.60 dB in peak signal-to-noise ratio and 0.072 in structural similarity; heterogeneity exceeded 99%, and clinical validation was limited.51

Model rankings differed by endpoint. In 51 extracted teeth with artificially degraded micro-CT, SwinIR had the highest peak signal-to-noise ratio, whereas a local texture estimator had the best perceptual scores and, in a two-method reader comparison with bicubic interpolation, significantly improved crack visibility (odds ratios, 1.326 and 1.310 at two thresholds).52 A confidence-aware panoramic transformer improved six image-quality metrics and reader opinion scores on 128 test images, but it did not test disease detection or calibrate its uncertainty map against errors.53 In 22 clinical maxillofacial CT cases, a semi-supervised spatial-frequency transformer improved reader-rated artifact suppression and intraoral-scan-referenced tooth integrity compared with supervised models.54 Conversely, in 48 extracted teeth with paired micro-CT, an edge-optimized adversarial model produced greater three-dimensional pulp/root-canal overlap than a hybrid attention transformer.55

CBCT translation and metal-artifact reduction

In the included technical and clinical studies, CBCT gray values varied according to device, field of view, exposure, and object position, and metal affected adjacent measurements or visibility.56,57,58,59,60,61 These studies did not test generative models.

Dental CBCT metal-artifact reduction was studied with convolutional and diffusion methods,49,62 whereas the spatial-frequency transformer study used maxillofacial CT rather than CBCT.54 Dental CBCT super-resolution or denoising studies used simulated or ex vivo material, extracted teeth, or technical image endpoints.52,55,63 The CBCT-to-CT studies identified for quantitative transfer involved brain, head-and-neck, chest, or lung radiotherapy cohorts rather than dental diagnosis.64,65,66,67

Within those declared transfer studies, a conditional denoising diffusion model reduced the mean absolute error of synthetic CT from 40.63 to 25.99 HU in brain cohorts and from 38.99 to 32.56 HU in head-and-neck cohorts.64 A lung diffusion model reduced the mean absolute error from 101.47 to 37.87 HU on institutional data but reached only 58.95 HU on a public dataset, a difference of approximately 21 HU between datasets.65 An unsupervised CycleGAN with vision-transformer encoder blocks was developed on head-and-neck and chest radiotherapy cohorts.66 In 19 head-and-neck patients with dental metallic implants, synthetic megavoltage CT reached formal dosimetric equivalence with the acquired reference, with mean Hounsfield unit differences between -1.1 and +9.0 HU and gamma pass rates of 94.3% and 97.6%.67

Accordingly, the radiotherapy studies provide evidence for Hounsfield unit or dosimetric endpoints, not for canal visibility, fracture detection, peri-implant assessment, or dental dimensional accuracy.64,65,66,67 Similarly, the dental metal-artifact review reported mainly image-quality endpoints and a high risk of bias in most studies.49

Structure and decision endpoints in generative studies

The biomedical super-resolution review reported invented or exaggerated image features.16 Within the dental studies, individual reports measured crack visibility, micro-CT-referenced root-canal structure, intraoral-scan-referenced tooth integrity, and blinded diagnosis or treatment planning.52,54,55,68 In the treatment-planning study, three clinicians assessed 302 CBCT scans from 90 patients. Enhancement improved mandibular-canal visibility for all readers and periodontal-ligament visibility for two readers, but it did not significantly change risk assessment, treatment strategy, or referral decisions.68

No study identified in this synthesis systematically counted the frequency or location of anatomy inserted or removed by a dental pixel-modifying model. The confidence-aware model supplied pixel-wise uncertainty but did not report whether uncertainty was calibrated to actual error.53 These findings are observed endpoint gaps in this corpus, not estimates of safety.

Table 2 summarizes the endpoints that were measured and the design limits associated with each evidence source.

Table 2. Evidence reported for dental and maxillofacial pixel-modifying models.

CBCT: cone-beam computed tomography, CT: computed tomography

Discussion

Across architecture-comparison reviews, transformer models frequently ranked first. A vision transformer led 58% of 21 dental comparisons in one review,5 and hybrid models consistently outperformed pure transformers in a 112-study review.6 The task-level primary evidence reproduced that pattern in some settings. A transformer ranked first for condyle segmentation,22 a global-local attention model retained an mIoU of 91.73% on an external intraoral-surface dataset,25 transformer landmark models were competitive on single-center benchmarks,38,39 and the forensic same-data comparison favored the hybrid model, with a mean absolute error of 2.61 years versus 3.08 years for the CNN and 7.48 years for the pure transformer.40

The same primary evidence also restricts the scope of those rankings. In the review reporting the 58% tally, 17 studies used private or institutional data, and 18 lacked external validation.5 In the largest matched segmentation comparison available here, two CNNs, two hybrid models, and two transformers were trained under the same five-fold scheme; CNNs were significantly better for teeth, tooth structures, and caries.21 Swin-Unet led on a private set but lost to CNNs on the public set,20 and both direct CBCT comparisons ranked a convolutional or attention-convolutional model above the tested transformers.23,24 Positive transformer results elsewhere did not isolate the architecture effect: the fracture detector with an AUC of 0.977 lacked a matched CNN backbone,27 and the condyle advantage was 0.047 Dice points across six models.22 Within this topic, the evidence supports complementarity in the literal sense that the leading architecture changed with task and dataset, but it counters use of the transformer label as a general performance claim.

Regarding external validation and transport across sites, the reviews and primary studies agree. External validation was absent from 18 of 21 studies in one review,5 the 112-study review estimated a 5%–8% internal-to-external performance decrease,6 and only four of the 36 source-checked comparisons in this synthesis used an independent external cohort. The external primary studies showed both retention (surface-segmentation mIoU, 92.14% to 91.73%)25 and degradation (lesion-detection F1 score, 0.926 to 0.781).29 Thus, the shared finding is scarcity of external evidence, whereas the direction of external change appears task-dependent. Because the 36-study set was not a systematic census of CNN comparators, the consistency of architecture ranking across all dental tasks remains untested.

The generative reviews converge by topic rather than contradicting one another. For field composition, the 43-study dental and maxillofacial review found that image-quality work predominated, which is consistent with the extracted endpoints and indicates that the four panoramic studies in the primary search reflected under-retrieval.48 For technical performance, artifact and enhancement metrics generally improved, and the CBCT meta-analysis estimated gains of 3.60 dB in peak signal-to-noise ratio and 0.072 in structural similarity.51 The cautions were also consistent: most metal-artifact studies had a high risk of bias; heterogeneity in the CBCT meta-analysis exceeded 99%;49,51 and 72% of dental GAN studies had a low risk of bias across QUADAS-2 domains, whereas only 17% had low applicability concern.50

The primary studies refine the conclusion of “limited clinical validation” rather than overturning it. Crack visibility improved in an ex vivo reader experiment,52 and a low-dose CBCT study measured diagnosis and treatment planning in 90 patients but found better visibility without significant changes in risk assessment, treatment strategy, or referral.68 Model rankings also depended on endpoint: SwinIR led peak signal-to-noise ratio, a local texture estimator led the perceptual metrics and was the only learned model included in the crack-visibility comparison,52 and an adversarial model exceeded a hybrid attention transformer for micro-CT-referenced three-dimensional root-canal agreement.55 For metal-artifact reduction, the closest primary evidence used maxillofacial CT rather than dental CBCT,54 so the review-level superiority claim for transformer and physics-informed models remains review-level in this synthesis.49 The evidence therefore supports generative image enhancement but does not equate image-quality gain, transformer use, and clinical decision benefit.

Table 3 organizes these relations by topic.

Table 3. Topic-level agreement and disagreement between previous reviews and the source-checked primary evidence.

Four future questions follow directly from the analyzed results. First, whether an architecture effect survives a matched design remains unresolved: 15 of 36 source-checked comparisons used a fair same-data comparator, while the largest matched segmentation study ranked CNNs first.21 Replicating that design across targets would test an observed source of disagreement rather than add another unmatched leaderboard.

Second, transport across sites and devices remains unresolved. Independent external validation occurred in four of 36 comparisons, was absent from 18 of 21 studies in one review, and produced a 5%–8% performance decrease in another.5,6 The external studies already show both retention (surface-segmentation mIoU, 92.14% to 91.73%) and degradation (lesion-detection F1 score, 0.926 to 0.781), so the evidence-derived question is which task, scanner, and population factors explain those different trajectories.25,29

Third, concordance between technical, reader, and decision endpoints remains unresolved. Fifteen of 36 comparisons reached a diagnostic, reader, or clinical endpoint. In generative studies, peak signal-to-noise and perceptual metrics selected different models, while crack visibility was tested for only one learned model against bicubic interpolation;52 enhanced CBCT visibility did not change treatment or referral decisions.68 In the hierarchical model of diagnostic-imaging efficacy, technical image quality is the lowest level, and improvement at that level does not automatically carry over to diagnostic accuracy, treatment decisions, or outcomes.69 Studies that measure these endpoints in the same cases can determine whether a technical gain propagates to a clinical decision.

Fourth, the frequency of anatomy inserted or removed by dental pixel-modifying models remains unmeasured in the identified studies. Micro-CT and intraoral-scan references demonstrate feasible structure-level comparisons,54,55 and one model already outputs pixel-wise uncertainty without reported calibration.53 Direct insertion/deletion counts and calibration against structure-level error are therefore testable extensions of methods already present in the evidence base, not assumptions about hypothetical systems.

Several limitations should be considered. The search used one bibliographic database, and primary-query recall was demonstrably incomplete. Topic-specific supplementary searches recovered missed image-formation, segmentation, detection, landmarking, and registration evidence and were used to refine the synthesis. Because the corpus was not rebuilt around those searches, Figure 1, the provenance counts, and the 36-study source-checked set describe only the original 159 records. Other topic-specific gaps may remain.

Screening was performed by one reviewer; CNN comparators were purposefully rather than systematically selected; no formal risk-of-bias or certainty instrument was applied; and four comparative assessments remained abstract-based. Architecture comparisons are therefore exploratory, and absence from this corpus is not proof that evidence does not exist.

The support, qualification, and counterevidence categories are interpretive and were not graded with a formal certainty framework. Prior-review conclusions were compared only with primary studies retrieved and verified for this narrative synthesis; therefore, disagreement identifies a boundary or counterexample rather than a pooled contrary effect. The four future questions are limited to gaps directly observed in these records and are not practice guidelines or adoption recommendations.

In conclusion, the primary evidence supports previous reviews’ conclusion that CNNs and transformers are complementary, but it qualifies the reported tendency toward transformer superiority. CNNs won the six-model matched panoramic segmentation comparison and both direct CBCT comparisons in this synthesis; transformers led condyle segmentation and an externally tested intraoral-surface task. The leading architecture therefore changed with target, dataset, and validation design.20,21,22,23,24,25

Generative reviews establish technical image-quality gains but also high heterogeneity, a high risk of bias in dental metal-artifact studies, and limited clinical validation.48,49,50,51 Primary studies add reader and structural endpoints, but image-quality metrics selected different leading models, crack visibility was tested for only one learned model against interpolation, and improved low-dose CBCT visibility did not significantly change treatment or referral decisions.52,55,68

Future questions justified by this evidence include whether architecture effects persist under matched training, whether performance is transferable to independent sites and devices, whether technical gains propagate to diagnostic decisions, and how often pixel-modifying models insert or remove anatomy. The reviewed studies do not yet answer those four questions.

Data availability

Search strategies, record-level screening logs, structured extraction records, provenance and study-design audits, and source-verification tables are described in the Supplementary Material and are available from the corresponding author.