PeriaTrace: Tooth-Level Grounding Diagnosis and Patient-Facing Explanations from Periapical Radiographs
Authors: Zhenhuan Zhou, Yuchen Zhang, Jingbo Zhu, Shiqing Ma, Yanlin Wu, Tao Li
Organizations: College of Computer Science, Nankai University, Tianjin, 300350, China · Key Laboratory of Data and Intelligent System Security, Ministry of Education, Tianjin, 300350, China · Department of stomatology, Tianjin Union Medical Center, The First Affiliated Hospital of Nankai University, China, Tianjin, 300121, China · The College of Software of Nankai University, Tianjin, 300350, China · Department of Stomatology, The Second Hospital of Tianjin Medical University, Tianjin, 300211, China
Periapical radiographs (PRs) are central to endodontic diagnosis, but interpreting them and preparing reports are time-consuming. Technical language can limit patients' understanding, while existing automated methods often lack tooth-level links between findings and image evidence. We present PeriaTrace, a three-stage framework integrating tooth-level diagnosis, visually grounded report generation, and patient-friendly explanations and care recommendations. PeriaTrace was developed and evaluated using 12,877 PRs from an internal dataset and two external clinical cohorts. In internal evaluation, it achieved a macro Dice coefficient of 85.35%, BLEU-4 of 0.6504, an F1 score of 94.37% for joint tooth localization and Fédération Dentaire Internationale (FDI) identification, and a grounding-tag F1 score of 77.93%. In the two external cohorts, senior endodontists considered 91.16% and 84.52% of reports suitable for retention or revision without complete rewriting. Generation averaged 4.41 seconds per PR, several tens of times faster than manual drafting. Among 196 survey participants without medical backgrounds, 96.4% preferred reports with plain-language explanations and care recommendations, and 84.7% found visual localization cues helpful. PeriaTrace may enable faster, evidence-traceable reporting while supporting clinical review and patient communication. (Abridged version)
Figures & tables
Figure 1: Clinical workflow and computational pipeline of PeriaTrace. The framework comprises three stages: multitask visual prior extraction, localization-conditioned diagnostic reasoning using a Low-Rank Adaptation (LoRA Hu et al. 2021 )-adapted Qwen3-VL-8B model Bai et al. 2025 , and evidence-grounded report generation. PeriaTrace produces structured tooth-level diagnostic reports with machine-readable visual evidence references, patient-friendly explanations, and care recommendations from periapical radiographs.
Figure 2: Qualitative and quantitative performance of PeriaTrace in visual prior extraction. (A) Representative tooth localization and caries detection results. (B) Semantic segmentation results. (C1–C2) Class-specific performance for tooth localization and FDI identification and semantic segmentation, respectively.
A. Prior extraction and evidence grounding performance of PeriaTrace
Task
Metric
Performance
QCfor absolute FDI identification
F1 / Sensitivity / Specificity
74.18% / 60.36% / 99.53%
Tooth localization and identification
Localization F1 / Joint localization–FDI F1
97.66% / 94.37%
Caries localization
BBox F1 at IoU = 0.3
92.17%
Semantic segmentation
Macro Dice / Macro IoU / Pixel accuracy
85.35% / 75.40% / 93.86%
Report grounding
Grounding-tag F1 / Tooth-box / Lesion-bbox / Mask
77.93% / 86.53% / 71.27% / 68.43%
Table 1: Quantitative results on the internal test cohort.
Figure 3: Representative report generation results with reliable FDI tooth identification. Two cases illustrate raw PRs, bounding-box and mask grounding, and comparisons between expert reference reports and PeriaTrace-generated reports.
Figure 4: Representative report generation results with indeterminate FDI tooth identification. Two cases demonstrate relative tooth localization, visual evidence grounding, and comparisons between expert reference reports and PeriaTrace-generated reports.
Professional report, Q2 A/B (no or minor revision)
1081/1742 (62.06%)
1113/1742 (63.89%)
1054/1742 (60.51%)
451/943 (47.83%)
474/943 (50.27%)
428/943 (45.39%)
Professional report, Q2 A/B/C (retained without rewrite)
1603/1742 (92.02%)
1608/1742 (92.31%)
1588/1742 (91.16%)
813/943 (86.21%)
816/943 (86.53%)
797/943 (84.52%)
Patient explanation, Q3 A/B (direct use or minor revision)
1304/1742 (74.86%)
1132/1742 (64.98%)
1101/1742 (63.20%)
519/943 (55.04%)
462/943 (48.99%)
436/943 (46.24%)
Table 2: External multi-center clinician evaluation of PeriaTrace outputs.
A. Professional versus patient-friendly reports
Case
Outcome
Patient-friendly
Professional
Effect size
Holm P
1
Objective disease comprehension, n/N (%)
79/87 (90.8)
45/109 (41.3)
OR 14.04 (6.18–31.92)
<0.001
1
Objective actionability, n/N (%)
81/87 (93.1)
42/109 (38.5)
OR 21.54 (8.63–53.75)
<0.001
1
Perceived understandability, mean ± SD [Top-2, %]
4.02 ± 1.15 [74.7]
3.21 ± 1.25 [47.7]
d=0.67
<0.001
1
Perceived action clarity, mean ± SD [Top-2, %]
3.94 ± 1.15 [78.2]
2.71 ± 1.22 [23.9]
d=1.04
<0.001
1
Communication helpfulness, mean ± SD [Top-2, %]
4.09 ± 1.05 [82.8]
3.03 ± 1.37 [45.0]
d=0.86
<0.001
Table 3: Patient-centered evaluation of report comprehension, actionability, visual grounding, and presentation preferences in 196 non-medical participants.
Figure 5: Detailed architecture and processing workflow of PeriaTrace. (a) Multitask visual prior extraction using a shared SAM ViT-B encoder. (b) Geometric evidence binding, localization-conditioned diagnosis with Qwen3-VL, and evidence-grounded report generation. (c–e) Detailed architectures of the localization, segmentation, and quality-control heads, respectively.
Metric
Definition
Metric
Definition
A. Detection
Bounding-box IoU
IoU(bp,bg)=∣bp∪bg∣∣bp∩bg∣
Precision
Precision=TP+FPTP
Recall
Recall=TP+FNTP
F1
F1=2TP+FP+FN2TP
B. Semantic segmentation
Class-specific Dice
Dicek=∣Ak∣+∣Gk∣2∣Ak∩Gk∣
Class-specific IoU
IoUk=∣Ak∪Gk∣∣Ak∩Gk∣
MacroDice
MacroDice=Ks−11k=1∑Ks−1Dicek
MacroIoU
MacroIoU=Ks−11k=1∑Ks−1IoUk
Table 4: Definitions of quantitative evaluation metrics and statistical measures.
Accurate evaluation of multimodal large language models in dental panoramic radiography is limited by the lack of fine-grained, clinically reliable benchmarks that reflect expert interpretation. This work introduces PanDent, a large-scale benchmark built on expert-validated tooth-level annotations. It comprises 9,554 panoramic radiographs with comprehensive structured findings produced by experienced dentists and further validated by an oral and maxillofacial radiologist. Clinician-guided reports are constructed from these findings using clinician-defined reporting logic, establishing explicit correspondence between structured clinical evidence and free-text descriptions. The dataset supports complementary evaluation of report-level textual agreement and tooth-level clinical consistency, with 9,049 public-source cases for supervised training and an independent 505-case in-house benchmark. Experiments cover proprietary, general-domain open-source, and medical-specific models. Current models generate fluent reports but exhibit substantial errors in fine-grained localization and tooth-level diagnosis. Supervised fine-tuning on PanDent markedly improves report alignment and diagnostic accuracy, bringing predictions closer to expert dental interpretation. These results establish PanDent as a rigorous benchmark and training resource for clinically grounded dental multimodal reasoning. Dataset sample: https://huggingface.co/datasets/Desperado1103/Pandent_Sample; code: https://github.com/HKUDentistry/PanDent. Full data will be released upon publication under source redistribution terms.
Xiaohan Li, Xinyu Liu, Chang Liu +5
Faculty of Dentistry, The University of Hong Kong, Hong Kong, China · Department of Computing, Imperial College London, London, United Kingdom · University of Science and Technology of China, Hefei, China +2
Dental caries and endodontic disease are among the most common health conditions worldwide, and intraoral periapical radiographs are central to their detection, treatment planning, and follow-up. Automated tooth-level classification of these images, however, lacks reproducible benchmarks, is often evaluated with image-level splits that leak patients between training and test, and is rarely validated across clinics. This paper presents the first patient-level, leakage-aware classification benchmark for single-tooth intraoral periapical radiographs on the DentIRO dataset, which comprises 5,300 images from 3,243 patients across two clinics and four classes: Healthy, Caries, Crowned, and Root Canal. Five transfer-learning models are compared with patient-grouped stratified cross-validation, so that every patient remains within a single fold. DenseNet121 gave the strongest and most stable result at a mean macro-F1 of 0.9787, while the four ImageNet-initialized backbones performed comparably. A controlled comparison on a fixed architecture showed that chest-radiograph pretraining transferred less effectively than ImageNet initialization. Bidirectional cross-center validation produced a small average generalization gap of 0.0077, and Grad-CAM confirmed that predictions rely on clinically meaningful tooth regions rather than acquisition artifacts. The benchmark offers a rigorous and reproducible baseline for intraoral radiograph classification.
Md Jubaer Rahman, Ulas Bagci
Department of Information and Communication Technology, Islamic University, Kushtia, Bangladesh · Hybrid and Machine Intelligence Lab, Northwestern University, Chicago, IL, USA
Automatic tooth segmentation and FDI numbering on panoramic radiographs underpins computer-assisted dental diagnosis, yet which factors govern performance remains unclear. We assemble a corpus of 1,422 panoramic radiographs containing 42,142 expert-delineated tooth polygons across the 32-class FDI taxonomy, annotated by 30 dental practitioners and independently reviewed by two others, and use it to isolate input resolution, architecture and anatomical priors under a single evaluation protocol. First, resolution dominates: across a controlled 640/1024/1280 ablation, mask mAP50-95 rises 0.656 -> 0.710 -> 0.717 while mAP50 stays flat at ~0.982. Both gains are significant under a paired bootstrap over images (p < 0.001, p = 0.024); neither mAP50 change is distinguishable from zero. Added resolution buys boundary precision, not detection. Second, architecture is nearly irrelevant in-domain: a query-based transformer with 2.1x the parameters is statistically equivalent to a one-stage detector (95% CI [-0.0064, +0.0064]), only marginally better under domain shift, 5.5x slower on CPU and not executable under standard ONNX runtimes. Third, three targeted interventions fail: a LoRA-adapted self-supervised encoder underperforms, a promptable foundation segmenter degrades masks by 39%, and globally optimal anatomical label assignment yields +0.0007 despite correcting a constraint violated in 40% of out-of-domain predictions. Zero-shot transfer to an independent multi-centre cohort, verified overlap-free, costs 62% of mask mAP50-95 but only 18% of mAP50, reproducing the dissociation. Decomposing masks along the tooth axis localises the residual error to the apical third. Boundary precision is therefore the binding constraint, and effort is better directed at resolution and acquisition diversity than at architectural novelty.