PerioXrays: Advances in Dental Radiograph Analysis

• PerioXrays is a domain where dental radiographs are reinterpreted as structured data, enabling quantitative analysis of periodontal and periapical diseases.
• It spans periapical and panoramic modalities to support tasks such as lesion detection, retrieval, segmentation, and 3D reconstruction of oral structures.
• Advanced annotation strategies and hybrid computational methods yield robust metrics for disease staging and bone-loss analysis, enhancing clinical decision support.

PerioXrays denotes a family of dental radiographic uses centered on periodontal, periapical, and apical disease analysis. In recent literature, the term is used in more than one sense: it can refer to periodontal/periapical radiographs used for multimodal retrieval in Dental CLAIRES (Kabir et al., 2023), and it can also name a large-scale panoramic benchmark for apical periodontitis detection introduced alongside PerioDet (Fang et al., 25 Jul 2025). More broadly, the same research space includes tooth-wise periodontitis stage grading from periapical radiographs (Kabir et al., 2021), automated alveolar bone-loss severity and pattern analysis from intraoral periapical radiographs (Wimalasiri et al., 25 Jun 2025), keypoint-based periodontal bone-loss analysis (Banks et al., 5 Mar 2025), and large-scale periapical-radiograph segmentation resources such as PRAD-10K (Zhou et al., 10 Apr 2025). Taken together, these works define PerioXrays as an emerging technical domain in which dental X-rays are treated not merely as visual records, but as structured computational objects for retrieval, segmentation, geometric measurement, detection, and clinical decision support.

1. Semantic scope and imaging modalities

In the periodontal and endodontic literature represented here, PerioXrays spans both periapical radiographs and panoramic radiographs, with different tasks attached to each modality. Periapical radiographs are repeatedly treated as central for local disease analysis: Dental CLAIRES states that periapical radiographs are a “gold standard” for diagnosing periodontitis because they capture detailed anatomical structure and bony defects and are available in almost all dental clinics (Kabir et al., 2023). PRAD-10K similarly describes periapical radiographs as the most extensively utilized imaging modality in endodontics and periodontics because they capture detailed local lesions at low cost (Zhou et al., 10 Apr 2025). Intraoral periapical radiographs are also the exclusive modality in the DenPAR study for automated alveolar bone-loss severity and pattern analysis (Wimalasiri et al., 25 Jun 2025), and in the YOLOv8-pose keypoint pipeline for periodontal bone-loss analysis (Banks et al., 5 Mar 2025).

Panoramic radiographs occupy a complementary role. In PerioDet, PerioXrays is the panoramic benchmark dataset itself, comprising annotated orthopantomograms for apical periodontitis detection (Fang et al., 25 Jul 2025). Panoramic images are also the input modality for several 2D-to-3D reconstruction studies, including 3DPX and PX2Tooth, which aim to recover 3D oral structure from a single panoramic X-ray (Li et al., 2024, Ma et al., 2024). This suggests that, within current usage, PerioXrays is not a single imaging protocol but a modality-specific research program spanning local periapical diagnosis, panoramic lesion detection, and even CT-like or point-cloud reconstruction from 2D radiographs.

The scope is broader than conventional periodontal staging alone. Some works focus on radiographic bone loss and periodontitis staging (Kabir et al., 2021, Wimalasiri et al., 25 Jun 2025), some on apical periodontitis (Fang et al., 25 Jul 2025, Misra et al., 2018), some on semantic retrieval over periodontal descriptions (Kabir et al., 2023), and some on multi-class segmentation of teeth, bone, pulp, restorations, implants, and apical lesions (Zhou et al., 10 Apr 2025, Mashayekhi et al., 2023). A specialized extension appears in multi-energy X-ray projection imaging of metal oxide particles inside gingival tissue, where the target is foreign-body gingivitis and related peri-implant disease rather than tooth or bone segmentation in routine radiographs (Cortez et al., 2023).

2. Data resources and annotation regimes

Recent PerioXrays research is distinguished by a shift from small, manually curated collections toward larger benchmark datasets with more explicit annotation protocols. The resources below illustrate the main dataset regimes currently represented in the literature.

Resource	Modality and scale	Primary annotation or task
Dental CLAIRES	687 periapical radiographs from 45 adult periodontitis patients	Text–image retrieval with stage, region, age, gender, ethnicity (Kabir et al., 2023)
HYNETS dataset	700 periapical X-rays	Bone, tooth, CEJ masks and tooth-level stage labels (Kabir et al., 2021)
DenPAR	1,000 IOPA radiographs	Tooth masks, keypoints, bone level lines, severity and pattern labels (Wimalasiri et al., 25 Jun 2025)
PerioXrays benchmark	3,673 panoramic X-rays, 5,662 apical periodontitis instances	Bounding-box detection of apical periodontitis (Fang et al., 25 Jul 2025)
PRAD-10K	10,000 clinical periapical radiographs	Pixel-level labels for nine structures, lesions, and devices (Zhou et al., 10 Apr 2025)

The annotation strategies differ substantially across tasks. Dental CLAIRES pairs each radiograph with structured textual metadata built from periodontal stage, anatomical region, and demographics; three examiners independently annotated stage and region, and conflicts in staging were resolved by majority voting (Kabir et al., 2023). HYNETS similarly relies on three examiners for tooth-level periodontitis stage assignment and uses majority vote as the reference label, while separately annotating bone area, tooth, and CEJ masks (Kabir et al., 2021). DenPAR adds a denser geometric layer: tooth segmentation masks, CEJ keypoints, crest-intersection keypoints, apex points, alveolar bone level lines, and 720 expert pattern annotations for horizontal versus angular bone loss (Wimalasiri et al., 25 Jun 2025).

The large-scale datasets emphasize benchmark standardization. PerioDet’s PerioXrays dataset is patient-split into 3,000 training images and 673 test images, with a multi-stage review by four experienced professional dentists to ensure label accuracy and reliability (Fang et al., 25 Jul 2025). PRAD-10K uses two experienced endodontists plus a computer researcher, with cross-review between endodontists and a final consistency pass; it provides nine pixel-level labels—Tooth, Bone, Pulp, Root Canal Filling, Denture Crown, Dental Fillings, Implant, Orthodontic Devices, and Apical Periodontitis—and also includes image-level classification labels for periodontitis, apical periodontitis, and inadequate root canal fillings (Zhou et al., 10 Apr 2025).

These annotation regimes reveal an important structural distinction. Some PerioXrays resources are conceptual and multimodal, with text paired to images (Kabir et al., 2023); some are geometric, with keypoints and lines (Wimalasiri et al., 25 Jun 2025, Banks et al., 5 Mar 2025); some are dense segmentation benchmarks (Zhou et al., 10 Apr 2025, Mashayekhi et al., 2023); and some are small-target detection benchmarks in panoramic images (Fang et al., 25 Jul 2025). This suggests that PerioXrays is best understood as a layered annotation ecosystem rather than a single dataset format.

3. Core computational paradigms

A defining feature of PerioXrays research is methodological heterogeneity. The field includes contrastive multimodal retrieval, end-to-end segmentation–classification systems, keypoint-and-geometry pipelines, lesion detectors for panoramic radiographs, and transformer-style semantic segmentation.

Dental CLAIRES formulates periodontal radiograph search as contrastive language–image retrieval. It uses DistilBERT as the text encoder, producing a 768-dimensional vector, and ResNet-50 as the image encoder, producing a 2048-dimensional vector; both are projected into a shared 256-dimensional embedding space where cosine similarity is computed (Kabir et al., 2023). The system embeds text and image pairs as

$$\text{ET} = \text{TP}(\text{ITE}(A_i)), \qquad \text{EI} = \text{IP}(\text{IIE}(B_i)),$$

and learns alignment by maximizing similarity for true pairs and minimizing it for random pairs. Its targets are defined by

$$\text{targets} = \text{softmax}(-\text{TS} + \text{IS}),$$

with text and image losses each implemented as binary cross-entropy against these targets (Kabir et al., 2023).

HYNETS takes a different route: an end-to-end entangled segmentation and classification CNN for tooth-wise stage grading from periapical radiographs. Its pipeline combines three U-Net-based segmentation subnetworks—for bone area, tooth, and CEJ—with a fully convolutional tooth classifier. The total loss is the sum of the three segmentation losses and the classification loss, so segmentation and classification are jointly fine-tuned rather than frozen as a sequential pipeline (Kabir et al., 2021). An interpretable RBL computation layer then overlays the masks, extracts tooth contours, computes CEJ–bone and CEJ–root distances, and estimates bone loss percentage using

$$\text{Bone loss percentage} = \max\left( \frac{\text{Line 1L}}{\text{Line 2L}}, \frac{\text{Line 1R}}{\text{Line 2R}} \right)\times 100.$$

A second major paradigm is keypoint-driven geometric measurement. The DenPAR framework uses YOLOv8x for tooth detection, three separate Keypoint R-CNN models for CEJ, crest–tooth intersection, and apex points, and YOLOv8x-seg for tooth masks and bone-level masks (Wimalasiri et al., 25 Jun 2025). Bone-loss severity is computed as

$$\text{ABL}(\%) = \frac{d_{\text{CEJ–crest}}}{d_{\text{CEJ–apex}}}\times 100\%,$$

after projecting the three detected points onto a min-max line. Bone-loss pattern is then defined geometrically: tangents are constructed on the tooth face and crest line, and the angle

$\theta = \cos^{-1}(u \cdot v)$

is thresholded at $54.1372^\circ$ to distinguish angular from horizontal bone loss (Wimalasiri et al., 25 Jun 2025). The related YOLOv8-pose study also uses a keypoint-first formulation, but augments it with a heuristic post-processing module that snaps CEJ, BL, and RL keypoints onto the nearest edge pixels of a matched tooth segmentation mask (Banks et al., 5 Mar 2025).

Large-benchmark detection work emphasizes small-target object detection. PerioDet introduces Background-Denoising Attention and IoU-Dynamic Calibration for panoramic apical periodontitis detection (Fang et al., 25 Jul 2025). BDA refines each FPN feature map using

$P_i^{BD} = (1 + z_i)\cdot P_i \odot S_i,$

where $z_i$ is a channel-importance vector and $S_i$ is a similarity map derived from local feature projection and a global scene embedding. IDC replaces fixed IoU-thresholding with

$$T_{rp} = \max\left(0.25,\; 0.2 + 0.15\cdot\left(\frac{\sqrt{w\cdot h}}{A_s}\right)^\lambda\right),$$

and a dynamic IoU assignment rule based on anchor and regressed overlaps (Fang et al., 25 Jul 2025).

Finally, semantic segmentation studies push PerioXrays toward richer multi-class scene parsing. PRNet combines Multi-scale Wavelet Convolution Network blocks, Global-local Feature Weighting Matrices, and Channel Fusion Attention to segment nine clinically relevant categories in periapical radiographs (Zhou et al., 10 Apr 2025). Radious uses a BEIT-Adapter plus Mask2Former pipeline across panoramic, periapical, and bitewing X-rays, targeting teeth, roots, pulp chamber, restorations, endodontics, implants, bone graft material, and several cystic lesions (Mashayekhi et al., 2023). Classical image-processing pipelines remain present as well: earlier IOPA studies rely on denoising, thresholding, watershed, superpixels, HOG, blob detectors, and edge operators such as Canny, Sobel, Scharr, Frangi, and Roberts to localize periapical abnormalities (Misra et al., 2018).

4. Quantification, staging, and reported performance

A central theme across PerioXrays work is the conversion of radiographic structure into explicit quantitative disease measures. Periodontitis staging is repeatedly tied to radiographic bone loss under the 2018 classification: Stage 1 or Stage I corresponds to $$\text{targets} = \text{softmax}(-\text{TS} + \text{IS}),$$0, Stage 2 or Stage II to $$\text{targets} = \text{softmax}(-\text{TS} + \text{IS}),$$1 or $$\text{targets} = \text{softmax}(-\text{TS} + \text{IS}),$$2, and Stage 3 or Stage III to $$\text{targets} = \text{softmax}(-\text{TS} + \text{IS}),$$3, extending into the middle third or beyond (Kabir et al., 2023, Kabir et al., 2021). This staging logic underlies both retrieval captions in Dental CLAIRES and tooth-wise severity grading in HYNETS.

The reported retrieval results in Dental CLAIRES are strong for concept-level search. With image and text augmentation, the full model achieved Hit@1 $$\text{targets} = \text{softmax}(-\text{TS} + \text{IS}),$$4, Hit@2 $$\text{targets} = \text{softmax}(-\text{TS} + \text{IS}),$$5, Hit@3 $$\text{targets} = \text{softmax}(-\text{TS} + \text{IS}),$$6, Precision@1 $$\text{targets} = \text{softmax}(-\text{TS} + \text{IS}),$$7, Precision@3 $$\text{targets} = \text{softmax}(-\text{TS} + \text{IS}),$$8, and MRR $$\text{targets} = \text{softmax}(-\text{TS} + \text{IS}),$$9 (Kabir et al., 2023). Stratified by query specificity, it maintained Hit@3 $$\text{Bone loss percentage} = \max\left( \frac{\text{Line 1L}}{\text{Line 2L}}, \frac{\text{Line 1R}}{\text{Line 2R}} \right)\times 100.$$0 for both low-difficulty and hard queries, the latter combining diagnosis, region, and demographics. These results indicate that detailed natural-language queries can retrieve clinically aligned periapical radiographs from a curated repository.

HYNETS frames the task as interpretable CAD rather than retrieval and reports both segmentation and stage-assignment performance. On heatmap images, bone area segmentation reached DSC $$\text{Bone loss percentage} = \max\left( \frac{\text{Line 1L}}{\text{Line 2L}}, \frac{\text{Line 1R}}{\text{Line 2R}} \right)\times 100.$$1, tooth segmentation DSC $$\text{Bone loss percentage} = \max\left( \frac{\text{Line 1L}}{\text{Line 2L}}, \frac{\text{Line 1R}}{\text{Line 2R}} \right)\times 100.$$2, and CEJ line segmentation DSC $$\text{Bone loss percentage} = \max\left( \frac{\text{Line 1L}}{\text{Line 2L}}, \frac{\text{Line 1R}}{\text{Line 2R}} \right)\times 100.$$3; the corresponding Jaccard Index values were $$\text{Bone loss percentage} = \max\left( \frac{\text{Line 1L}}{\text{Line 2L}}, \frac{\text{Line 1R}}{\text{Line 2R}} \right)\times 100.$$4, $$\text{Bone loss percentage} = \max\left( \frac{\text{Line 1L}}{\text{Line 2L}}, \frac{\text{Line 1R}}{\text{Line 2R}} \right)\times 100.$$5, and $$\text{Bone loss percentage} = \max\left( \frac{\text{Line 1L}}{\text{Line 2L}}, \frac{\text{Line 1R}}{\text{Line 2R}} \right)\times 100.$$6, respectively (Kabir et al., 2021). For stage assignment, Figure 6a reports AUC $$\text{Bone loss percentage} = \max\left( \frac{\text{Line 1L}}{\text{Line 2L}}, \frac{\text{Line 1R}}{\text{Line 2R}} \right)\times 100.$$7 for Stage I, $$\text{Bone loss percentage} = \max\left( \frac{\text{Line 1L}}{\text{Line 2L}}, \frac{\text{Line 1R}}{\text{Line 2R}} \right)\times 100.$$8 for Stage II, and $$\text{Bone loss percentage} = \max\left( \frac{\text{Line 1L}}{\text{Line 2L}}, \frac{\text{Line 1R}}{\text{Line 2R}} \right)\times 100.$$9 for Stage III, with average AUC $$\text{ABL}(\%) = \frac{d_{\text{CEJ–crest}}}{d_{\text{CEJ–apex}}}\times 100\%,$$0. A rules-based baseline using segmentation-derived RBL alone achieved AUC $$\text{ABL}(\%) = \frac{d_{\text{CEJ–crest}}}{d_{\text{CEJ–apex}}}\times 100\%,$$1, indicating that the entangled classifier adds substantial value. Agreement analysis further showed HYNETS versus professor $$\text{ABL}(\%) = \frac{d_{\text{CEJ–crest}}}{d_{\text{CEJ–apex}}}\times 100\%,$$2, and Student’s t-test comparing HYNETS-measured RBL with expert measurements yielded $$\text{ABL}(\%) = \frac{d_{\text{CEJ–crest}}}{d_{\text{CEJ–apex}}}\times 100\%,$$3, indicating no significant difference (Kabir et al., 2021).

The DenPAR framework quantifies both severity and defect morphology. Its severity estimates achieved ICC $$\text{ABL}(\%) = \frac{d_{\text{CEJ–crest}}}{d_{\text{CEJ–apex}}}\times 100\%,$$4 on training, $$\text{ABL}(\%) = \frac{d_{\text{CEJ–crest}}}{d_{\text{CEJ–apex}}}\times 100\%,$$5 on validation, and $$\text{ABL}(\%) = \frac{d_{\text{CEJ–crest}}}{d_{\text{CEJ–apex}}}\times 100\%,$$6 on test, which the paper interprets using Koo and Li’s guidelines as good reliability (Wimalasiri et al., 25 Jun 2025). For pattern classification, 81 of 720 sites were excluded due to missing bone lines or masks, leaving 639 evaluated cases; accuracy was $$\text{ABL}(\%) = \frac{d_{\text{CEJ–crest}}}{d_{\text{CEJ–apex}}}\times 100\%,$$7, precision $$\text{ABL}(\%) = \frac{d_{\text{CEJ–crest}}}{d_{\text{CEJ–apex}}}\times 100\%,$$8, recall $$\text{ABL}(\%) = \frac{d_{\text{CEJ–crest}}}{d_{\text{CEJ–apex}}}\times 100\%,$$9, and sensitivity $\theta = \cos^{-1}(u \cdot v)$0. The same study reports YOLOv8x tooth detection precision $\theta = \cos^{-1}(u \cdot v)$1, mAP50 $\theta = \cos^{-1}(u \cdot v)$2, and mAP50:95 $\theta = \cos^{-1}(u \cdot v)$3, as well as Keypoint R-CNN superiority over YOLOv8 Pose across CEJ, crest-intersection, and apex keypoints (Wimalasiri et al., 25 Jun 2025).

The YOLOv8-pose periodontal bone-loss study emphasizes keypoint metrics. With heuristic post-processing, it achieved PRCK $\theta = \cos^{-1}(u \cdot v)$4, PRCK $\theta = \cos^{-1}(u \cdot v)$5, mAP $\theta = \cos^{-1}(u \cdot v)$6 for tooth object detection, mesial dice score $\theta = \cos^{-1}(u \cdot v)$7 for periodontal staging, and dice score $\theta = \cos^{-1}(u \cdot v)$8 for furcation involvement (Banks et al., 5 Mar 2025). This is materially weaker than DenPAR’s severity-and-pattern pipeline on its larger dataset, but it introduces a stage-agnostic keypoint schema and the PRCK metric, which normalizes keypoint error to average tooth size in the image.

For large-scale benchmark detection, PerioDet reports $\theta = \cos^{-1}(u \cdot v)$9, $54.1372^\circ$0, $54.1372^\circ$1, and $54.1372^\circ$2 on the PerioXrays panoramic dataset (Fang et al., 25 Jul 2025). Its human–computer collaborative experiment on 100 panoramic images further reports precision increasing from 73.1% to 92.5%, recall from 74.3% to 96.1%, and reading efficiency improving from about 28 s/image to about 13 s/image with PerioDet assistance. In the segmentation-benchmark regime, PRNet reaches an average DSC of 84.24% on PRAD-10K and records class-wise DSC values including 92.38 for Tooth, 93.02 for Bone, 88.87 for Pulp, 92.46 for Orthodontic Devices, and 88.83 for Apical Periodontitis (Zhou et al., 10 Apr 2025).

5. Research and clinical workflows

PerioXrays systems are increasingly designed not only as offline models but as interfaces for active research workflows. Dental CLAIRES explicitly provides a GUI with a query input box, a control for the number of images to return, and an output panel displaying retrieved periapical radiographs (Kabir et al., 2023). This supports dataset exploration by concept, case-based teaching set construction, model validation through interactive error analysis, and semi-automatic image pre-filtering for downstream studies such as bone-loss quantification or progression analysis.

HYNETS also includes a web interface tailored to radiographic decision support. A clinician uploads a periapical radiograph, and the system returns segmentation overlays for bone area, tooth contours, and CEJ line, as well as per-tooth CEJ–bone distance, CEJ–root distance, bone loss percentage, and final stage assignment (Kabir et al., 2021). Because the RBL computation is geometric and inspectable, the interface is not purely classificatory; it attempts to preserve conventional periodontal reasoning inside the prediction pipeline.

The DenPAR study presents a workflow closer to chairside automation. A periapical radiograph is captured as usual; the system detects each tooth, computes a numeric percentage of bone loss per tooth side, flags horizontal versus angular defects, and can in principle overlay keypoints, lines, and color-coded outputs on the image (Wimalasiri et al., 25 Jun 2025). The paper positions this as support for decision-making, screening and triage, and longitudinal follow-up, where objective, reproducible measurements are preferable to subjective visual comparison.

PRAD-10K and PRNet broaden the workflow beyond a single disease endpoint. Because the nine-class segmentation includes tooth, bone, pulp, root canal filling, denture crown, dental fillings, implant, orthodontic devices, and apical periodontitis, a single model output can support lesion localization, treatment-quality assessment, restoration documentation, and structural context for endodontic or periodontal interpretation (Zhou et al., 10 Apr 2025). Radious extends this logic further by treating dental radiographs as multi-class semantic scenes across panoramic, periapical, and bitewing views, with outputs that include teeth, roots, endodontics, crowns, implants, bone graft material, and cystic lesions (Mashayekhi et al., 2023).

A more specialized workflow appears in the physics-oriented gingival micro-imaging study. There, the proposed use case is rapid, non-destructive screening of biopsies for foreign-body gingivitis or peri-implant disease using multi-energy X-ray projection imaging (Cortez et al., 2023). This is not a routine dental radiograph workflow, but it expands the PerioXrays concept toward bench-top micro-projection systems for soft-tissue foreign-particle detection.

6. Limitations, disagreements, and future directions

Several limitations recur across the literature. Many datasets are single-institution and moderate in size: Dental CLAIRES uses 687 periapical radiographs from one dental school (Kabir et al., 2023), HYNETS uses 700 periapical X-rays from one institution (Kabir et al., 2021), DenPAR uses 1,000 IOPA radiographs (Wimalasiri et al., 25 Jun 2025), and PRAD-10K, though much larger, is still derived from one top-tier hospital’s endodontics department (Zhou et al., 10 Apr 2025). PerioDet is multi-center, but its data come from a limited set of hospitals in one country (Fang et al., 25 Jul 2025). Generalizability across acquisition protocols, populations, vendors, and disease spectra therefore remains an open problem.

A second limitation is label uncertainty. Dental CLAIRES notes that Cohen’s Kappa between annotators can be as low as 0.23 for staging, and HYNETS reports fair-to-substantial inter-examiner agreement rather than perfect consensus (Kabir et al., 2023, Kabir et al., 2021). This complicates any strong claim that radiographic stage labels are fully objective ground truth. A plausible implication is that PerioXrays models should increasingly be assessed not only by overlap or AUC, but by agreement with expert consensus distributions and by robustness under label ambiguity.

Modality restriction is another structural constraint. Some systems are periapical-only (Kabir et al., 2021, Wimalasiri et al., 25 Jun 2025, Zhou et al., 10 Apr 2025), some panoramic-only (Fang et al., 25 Jul 2025), and some do not integrate CBCT or richer periodontal charting (Kabir et al., 2023, Wimalasiri et al., 25 Jun 2025). The DenPAR study, for example, quantifies severity and pattern from IOPA but does not incorporate probing depth, clinical attachment level, or furcation-focused 3D imaging (Wimalasiri et al., 25 Jun 2025). The YOLOv8-pose study explicitly reports weak performance on detached periodontal ligament space and furcation involvement (Banks et al., 5 Mar 2025). These gaps matter because much of periodontal diagnosis is multimodal by construction.

The future direction most consistently proposed is integration: more data types, more institutions, more modalities, and richer text or clinical metadata. Dental CLAIRES explicitly suggests adding other abnormalities such as caries, endodontic lesions, and implants; other imaging modalities such as panoramic radiographs and CBCT; and clinical data such as medical history, vital signs, and medication (Kabir et al., 2023). PRAD-10K points toward more efficient supervised models, semi-supervised learning, and multimodal PR analysis (Zhou et al., 10 Apr 2025). Radious explicitly identifies bone recession assessment for surgical planning as future work (Mashayekhi et al., 2023).

A more ambitious frontier is 3D reconstruction from 2D radiographs. 3DPX reconstructs a flattened 3D oral volume from a single panoramic X-ray using a progressive hybrid MLP-CNN network and improves downstream angular misalignment classification (Li et al., 2024). PX2Tooth reconstructs 3D point-cloud teeth from a single panoramic X-ray, reaching IoU $54.1372^\circ$3 and specifically improving the root apex region via a Prior Fusion Module (Ma et al., 2024). PerX2CT shows that two approximately perpendicular X-rays can be used for perspective-projection-based CT reconstruction in another anatomical domain, with coordinate-wise local and global feature sampling (Kyung et al., 2023). These studies do not yet deliver full periodontal 3D diagnosis, and PX2Tooth explicitly notes that jawbone is not reconstructed (Ma et al., 2024). Even so, they suggest that future PerioXrays systems may move from 2D measurement and lesion marking toward low-dose 3D structural inference.

In its current state, PerioXrays is therefore best viewed as an evolving computational radiology domain rather than a single platform. Its contemporary forms include multimodal retrieval over periapical radiographs, interpretable stage grading, geometric bone-loss measurement, large-scale panoramic lesion detection, dense periapical segmentation, and early attempts at 3D inference from 2D dental X-rays. The unifying idea is consistent across these variants: dental radiographs are being transformed from static diagnostic images into structured, queryable, and quantitatively analyzable representations of periodontal and periapical disease.