PerioDet: Clinical Apical Periodontitis Detection

Updated 7 July 2026

The paper introduces PerioDet, integrating Background-Denoising Attention and IoU-Dynamic Calibration to improve AP metrics by +5.8 over leading CNN baselines.
PerioDet leverages a multi-scale Feature Pyramid Network with size-aware, dynamic label assignment to effectively suppress background noise in noisy panoramic X-rays.
Clinical evaluations indicate that PerioDet enhances diagnostic efficiency by reducing image interpretation time and boosting the recall of small, subtle lesions.

Clinical-Oriented Apical Periodontitis Detection (PerioDet) is a detection paradigm for automated identification of apical periodontitis on panoramic dental radiographs. It was introduced together with the PerioXrays benchmark to address two panoramic-specific obstacles: substantial background noise in panoramic X-rays and the predominance of small, subtle lesions at root apices. Architecturally, PerioDet operates on a Feature Pyramid Network (FPN) representation and augments a conventional detector with Background-Denoising Attention (BDA) and IoU-Dynamic Calibration (IDC), thereby combining feature denoising with size-aware, training-progress-aware label assignment. Within the broader literature on apical periodontitis computer-aided diagnosis, it occupies a distinct position between earlier interpretable image-processing pipelines for intra-oral periapical radiographs, segmentation-oriented deep models for periapical radiographs, and tooth-level panoramic screening systems for periapical radiolucency (Fang et al., 25 Jul 2025, Misra et al., 2018, Zhou et al., 10 Apr 2025, Wang et al., 14 Feb 2025).

1. Clinical target and radiographic formulation

Apical periodontitis is a prevalent oral pathology and a persistent target for radiographic decision support. In intra-oral periapical radiographs, radiographic signs described in the literature include periapical radiolucency, widened periodontal ligament space, loss of lamina dura, and changes in trabecular pattern. An earlier IOPA case study operationalized apical periodontitis visually through periodontal ligament widening and periapical radiolucency, then used noise removal, edge detection, thresholding, and segmentation to make these signs more conspicuous and to localize regions of interest near tooth apices (Misra et al., 2018).

The panoramic setting is more difficult. PerioDet is motivated by the observation that panoramic X-rays exhibit dense anatomical structures, low contrast, and artifacts that obscure subtle apical lesions. A related multinational panoramic study operationalized apical periodontitis as “periapical radiolucency” at the tooth level and reported that human inter-reader agreement for this finding was the lowest across the evaluated findings on the Taiwan* subset, with Cohen’s kappa of 37.6% (95% CI: 14.7%–60.6%). This indicates that the target is not only visually subtle but also diagnostically variable in routine panoramic interpretation (Wang et al., 14 Feb 2025).

These observations clarify the clinical orientation of PerioDet. It is not merely a generic object detector applied to dental images; it is formulated around a lesion class whose radiographic expression is subtle, small-scale, and highly confounded by surrounding anatomy. A plausible implication is that model design for this task must attend simultaneously to foreground enhancement, clutter suppression, and positive-sample sufficiency for small lesions.

2. PerioXrays as benchmark infrastructure

PerioDet was introduced alongside PerioXrays, a large-scale panoramic radiograph benchmark comprising 3,673 images from 3,482 unique patients and 5,662 meticulously annotated instances of apical periodontitis. The images were collected between 2022 and 2024 from multiple hospitals, dental clinics, and orthodontic centers, then standardized to a resolution of 1333 × 800 pixels to ensure consistent input size while minimizing information loss. The paper describes PerioXrays as, to the best of the authors’ knowledge, the first benchmark dataset for automated apical periodontitis diagnosis (Fang et al., 25 Jul 2025).

Annotation is box-based rather than mask-based. Bounding boxes were created with a custom tool and adjusted to the true extent of the lesion, avoiding overextension or undersizing. Diagnostic criteria followed established clinical standards for apical pathology, citing Abbott (2004), Huumonen (2002), and Armitage (1995). Each image underwent a multi-stage review by four experienced professional dentists. At the same time, several dataset-governance details remain unreported: scanner models, bit-depth, file formats, inter-rater reliability statistics such as Cohen’s kappa, and explicit exclusion criteria are not provided in the paper (Fang et al., 25 Jul 2025).

The split is patient-level, with 3,000 training images and 673 test images. Distribution analysis uses an area ratio (AR) metric defined as the proportion of the image occupied by the lesion. The benchmark exhibits a high prevalence of small lesions, with many instances having $AR \leq 0.5\%$ . This distribution is central to the design of PerioDet, because conventional detectors can underperform when positive-anchor assignment becomes unstable for very small targets (Fang et al., 25 Jul 2025).

PerioXrays is modality-specific. It complements, rather than duplicates, other dental datasets. PRAD-10K, for example, contains 10,000 periapical radiograph images with pixel-level annotations for nine classes, including apical periodontitis, and supports segmentation rather than box detection on full periapical radiographs. By contrast, an older IOPA study was a single-case image-processing demonstration without train/validation/test splits or quantitative validation. This suggests that PerioXrays fills a distinct benchmark niche for panoramic, lesion-level detection under clinically realistic clutter conditions (Zhou et al., 10 Apr 2025, Misra et al., 2018).

3. Background-Denoising Attention

PerioDet is built on an FPN for multi-scale feature processing, with feature maps at different scales denoted $P_i$ . Its first task-specific module, Background-Denoising Attention, is designed to refine these features by suppressing clutter while preserving lesion evidence. The motivation is explicit: panoramic radiographs contain dense anatomy, low contrast, and artifacts, and these factors obscure subtle apical lesions (Fang et al., 25 Jul 2025).

For an FPN feature map $P_i$ , BDA generates a refined map $P_i^{BD}$ using a learned channel importance vector $z_i$ and a target-scene similarity map $S_i$ . The formulation is

$\tilde{P}_i = \text{ReLU}(\text{BN}(\text{Conv}_{1 \times 1}(P_i))),$

$u = \text{Conv}_{1 \times 1}(C_5),$

$S_i = \text{Sigmoid}(\tilde{P}_i \cdot u),$

and

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

Here, $P_i$ 0 is a scene embedding vector derived from the high-level feature map $P_i$ 1, and the interaction $P_i$ 2 produces a per-location gating signal after broadcasting as appropriate. The Hadamard product with $P_i$ 3 imposes context-aware spatial gating, while $P_i$ 4 applies channel reweighting. The paper specifies the intent of $P_i$ 5—to prioritize channels capturing rich spatial details and suppress background noise—but does not provide the explicit formula for $P_i$ 6 in the excerpt (Fang et al., 25 Jul 2025).

Conceptually, BDA functions as a denoising prior over the lesion-background mixture characteristic of panoramic dental imaging. It combines channel attention with scene-conditioned gating, anchoring local evidence to high-level context. No additional regularization loss for BDA is reported; the module is trained end-to-end within the underlying detection framework (Fang et al., 25 Jul 2025).

The role of BDA is easier to appreciate when contrasted with earlier handcrafted pipelines. In IOPA radiographs, interpretable clutter reduction was attempted with median filtering, bilateral filtering, morphological opening and closing, masked Otsu thresholding, SLIC superpixels, watershed segmentation, and region adjacency graphs. PerioDet replaces this sequence of manually staged preprocessing with a learned attention mechanism integrated directly into the detector backbone, while preserving the same broad clinical aim: reducing noise-driven false positives and emphasizing subtle periapical structures (Misra et al., 2018).

4. IoU-Dynamic Calibration and training protocol

The second task-specific module in PerioDet is IoU-Dynamic Calibration, which addresses the instability of IoU-based positive assignment for small apical lesions. The core problem is that minor localization errors can cause large IoU fluctuations when lesions occupy very few pixels, producing insufficient positive anchors and undertraining. IDC introduces size-aware adaptive thresholds and a dynamic label-assignment criterion that incorporates both anchor quality and current regression quality (Fang et al., 25 Jul 2025).

The baseline overlap measure is

$P_i$ 7

IDC defines an adaptive IoU threshold for positive samples as

$P_i$ 8

where $P_i$ 9 and $P_i$ 0 are the ground-truth box width and height, $P_i$ 1 is a dataset-specific minimal area reference, and $P_i$ 2 controls growth rate. Smaller boxes therefore receive lower effective thresholds, increasing the number of positive anchors (Fang et al., 25 Jul 2025).

The dynamic assignment score is

$P_i$ 3

where $P_i$ 4 is anchor-to-ground-truth IoU, $P_i$ 5 is regression-box-to-ground-truth IoU, and $P_i$ 6 are hyperparameters. This term penalizes disparity between anchor matching quality and regressed localization, promoting training stability. The scheduling of $P_i$ 7 over training progress $P_i$ 8 is

$P_i$ 9

Early training therefore depends entirely on anchor-ground-truth IoU, and later training progressively incorporates regression quality. In ablations, $P_i^{BD}$ 0, $P_i^{BD}$ 1, and $P_i^{BD}$ 2 yielded optimal performance (Fang et al., 25 Jul 2025).

The label-assignment procedure is explicit: for each anchor, compute $P_i^{BD}$ 3 from object size and $P_i^{BD}$ 4 from the current state; designate anchors with $P_i^{BD}$ 5 as positive. PerioDet is then trained end-to-end using the standard detection losses of the underlying detector, with IDC altering only positive/negative assignment. The reported training setup uses AdamW, batch size 2, a 1× schedule of 12 epochs, initial learning rate $P_i^{BD}$ 6, and weight decay $P_i^{BD}$ 7, with ResNet-50 as the backbone for all baselines. Batch Normalization is specified within BDA blocks, but augmentations, normalization strategy beyond that, pretraining, early stopping, NMS type, inference thresholds, runtime, and memory footprint are not reported (Fang et al., 25 Jul 2025).

5. Benchmark performance and ablation structure

Evaluation on PerioXrays is reported with AP overall, AP50, AP75, and AP for small, medium, and large objects. PerioDet achieved AP 53.5, AP50 84.2, AP75 55.6, $P_i^{BD}$ 8 42.3, $P_i^{BD}$ 9 50.9, and $z_i$ 0 58.7. Among CNN-based baselines, the strongest reported AP is CFINet at 47.7; among transformer-based baselines, the strongest is Salience-DETR at 50.3. On this basis, PerioDet improves over the best CNN baseline by +5.8 AP and over the best transformer baseline by +3.2 AP. The small-object result is especially prominent: $z_i$ 1, compared with the strongest baseline small-object AP in the 37.2–37.0 range (Fang et al., 25 Jul 2025).

The baseline set is heterogeneous. CNN detectors include Faster R-CNN, ATSS, YOLOX, Sparse R-CNN, RFLA, and CFINet. Transformer detectors include Deformable-DETR, Conditional-DETR, DAB-DETR, DINO, Co-DINO, and Salience-DETR. Their reported scores show that panoramic apical periodontitis detection is challenging even for recent detectors. DINO, for example, attains AP 48.9 but only $z_i$ 2 20.4, whereas Co-DINO and Salience-DETR recover stronger small-object performance at 36.8 and 37.0, respectively. This distribution indicates that overall AP alone is insufficient for this task; performance on small lesions is a central discriminator (Fang et al., 25 Jul 2025).

Ablation results isolate the contributions of the two task-specific modules. The base detector without BDA or IDC yields AP 49.9 and $z_i$ 3 36.8. Adding BDA alone gives AP 51.7 and $z_i$ 4 40.6; adding IDC alone gives AP 52.3 and $z_i$ 5 41.1; combining both yields AP 53.5 and $z_i$ 6 42.3. The reported interpretation is that BDA primarily reduces background-induced errors, while IDC elevates small-lesion recall and localization stability. Statistical significance tests are not reported, so the results establish empirical superiority within the benchmark but do not quantify uncertainty around the improvements (Fang et al., 25 Jul 2025).

Several common metrics remain absent from the benchmark report. The paper does not provide $z_i$ 7, precision, recall, F1, sensitivity, specificity, or ROC-AUC. This matters because PerioDet is clinically positioned, yet its algorithmic benchmark is expressed purely in detection AP metrics. A plausible implication is that the benchmark and the clinical study should be read together rather than separately.

PerioDet includes a human-computer collaborative experiment intended to assess clinical applicability. Six dentists participated—two junior, two intermediate, and two senior—and none had previously been exposed to the data. On a random set of 100 images from PerioXrays, each participant made two consecutive diagnostic judgments per image, first without assistance and then with PerioDet assistance. Reported outcomes were precision 73.1% to 92.5%, recall 74.3% to 96.1%, and efficiency from approximately 28 s/image to approximately 13 s/image. The paper interprets these gains as evidence that PerioDet can function as an auxiliary diagnostic tool for professional dentists, although statistical tests and detailed false-positive/false-negative analyses are not provided (Fang et al., 25 Jul 2025).

This clinical framing is consistent with a wider pattern in dental AI validation. A paired-reader intraoral radiograph study on apical lesions reported tooth-level sensitivity increasing from 70.4% to 90.7%, specificity decreasing from 98.7% to 97.2%, and apical-lesion AUC increasing from 0.70 to 0.92 when dentists used algorithmic assistance. Separately, a multinational panoramic system for periapical radiolucency achieved on the Taiwan* subset AI sensitivity 93.5%, specificity 97.0%, precision 61.1%, F1 73.9%, F2 84.5%, and AUC-ROC 97.4%, while average human readers achieved sensitivity 25.5%, specificity 99.5%, precision 73.1%, F1 37.0%, and F2 29.1%. That study also reported the AI processing images 79 times faster than human readers, with mean AI time per DPR of 1.55 s versus mean human reading time of 122 s (Leemput et al., 2024, Wang et al., 14 Feb 2025).

PerioDet should also be situated relative to modality-specific alternatives. PRAD-10K and PRNet target periapical radiographs rather than panoramic radiographs and treat apical periodontitis as a segmentation class. PRNet reported an apical periodontitis Dice score of 88.83%, the highest among the compared models on PRAD-10K. This does not directly compete with PerioDet’s box-detection setting, but it shows that the field is developing along at least two technical axes: panoramic detection under heavy scene clutter and periapical segmentation with pixel-level labels (Zhou et al., 10 Apr 2025).

Several limitations constrain interpretation. PerioXrays is multi-institutional, but scanner models, bit-depth, and other device metadata are not documented; inter-rater reliability statistics are not reported; and the benchmark is confined to panoramic radiographs, leaving applicability to CBCT or periapical intraoral films unresolved. The paper does not enumerate failure modes, though extremely subtle lesions, severe artifacts, and overlapping anatomical structures are identified as plausible challenges. It also does not report operating-point calibration, uncertainty estimation, or broader multi-center external validation with device harmonization. These omissions are consequential because related panoramic work has already shown site-dependent sensitivity variation despite robust AUC, and because weak human agreement for periapical radiolucency suggests that label definition itself remains a difficult clinical variable (Fang et al., 25 Jul 2025, Wang et al., 14 Feb 2025).

In this sense, PerioDet represents both a concrete detector and a benchmarked research agenda. Its substantive contribution is the explicit coupling of background-denoising attention and dynamic IoU-based calibration to the lesion statistics of panoramic apical periodontitis, together with a publicly accessible benchmark at https://github.com/XiaochengFang/MICCAI2025_PerioDet. Its unresolved questions concern generalizability, calibration, annotation reliability, and extension across imaging modalities.