List of Papers Browse by Subject Areas Author List
Abstract
Deep learning (DL), a pivotal technology in artificial intelligence, has recently gained substantial traction in the domain of dental auxiliary diagnosis. However, its application has predominantly been confined to imaging modalities such as panoramic radiographs and Cone Beam Computed Tomography, with limited focus on auxiliary analysis specifically targeting Periapical Radiographs (PR). PR are the most extensively utilized imaging modality in endodontics and periodontics due to their capability to capture detailed local lesions at a low cost. Nevertheless, challenges such as projection angle and artifacts complicate the annotation and recognition of PR, leading to a scarcity of publicly available, large-scale, high-quality PR analysis datasets. This scarcity has somewhat impeded the advancement of DL applications in PR analysis. In this paper, we present PRAD-10K, a dataset for PR analysis. PRAD-10K comprises 10,000 clinical periapical radiograph images, with pixel-level annotations provided by professional endodontists for nine distinct anatomical structures, lesions, and artificial restorations or medical devices. We also include classification labels for images with typical conditions or lesions. Furthermore, we introduce a DL network named PRNet to establish benchmarks for PR segmentation tasks. Experimental results demonstrate that PRNet surpasses previous state-of-the-art medical image segmentation models on the PRAD-10K dataset. The code and dataset will be released at https://github.com/nkicsl/PRAD.
Links to Paper and Supplementary Materials
Main Paper (Open Access Version): https://papers.miccai.org/miccai-2025/paper/0247_paper.pdf
SharedIt Link: Not yet available
SpringerLink (DOI): Not yet available
Supplementary Material: Not Submitted
Link to the Code Repository
https://github.com/nkicsl/PRAD
Link to the Dataset(s)
PRAD dataset: https://github.com/nkicsl/PRAD
BibTex
@InProceedings{ZhoZhe_PRAD_MICCAI2025,
author = { Zhou, Zhenhuan and Zhang, Yuchen and Xu, Ruihong and Zhao, Xuansen and Li, Tao},
title = { { PRAD: Periapical Radiograph Analysis Dataset and Benchmark Model Development } },
booktitle = {proceedings of Medical Image Computing and Computer Assisted Intervention -- MICCAI 2025},
year = {2025},
publisher = {Springer Nature Switzerland},
volume = {LNCS 15972},
month = {September},
page = {477 -- 486}
}
Reviews
Review #1
- Please describe the contribution of the paper
This paper introduce PRAD-10K, a PR dataset featuring expert annotations, which serves as a potential benchmark for research in DL based PR image analysis. Additionally, to tackle the multi-scale challenges inherent in PR image segmentation tasks, a DL network is used to integrate the Multi scale Wavelet Convolution Network (MWCN) and the Channel Fusion Attention mechanism.
- Please list the major strengths of the paper: you should highlight a novel formulation, an original way to use data, demonstration of clinical feasibility, a novel application, a particularly strong evaluation, or anything else that is a strong aspect of this work. Please provide details, for instance, if a method is novel, explain what aspect is novel and why this is interesting.
This paper provide a Periapical Radiograph dataset.But the novality of proposed model including the Multi-scale Wavelet Convolution Network (MWCN) and the Channel Fusion Attention is too weak.
- Please list the major weaknesses of the paper. Please provide details: for instance, if you state that a formulation, way of using data, demonstration of clinical feasibility, or application is not novel, then you must provide specific references to prior work.
The novality of proposed model including the Multi-scale Wavelet Convolution Network (MWCN) and the Channel Fusion Attention is too weak. Neither dataset processing nor parameters are described.
- Please rate the clarity and organization of this paper
Satisfactory
- Please comment on the reproducibility of the paper. Please be aware that providing code and data is a plus, but not a requirement for acceptance.
The authors claimed to release the source code and/or dataset upon acceptance of the submission.
- Optional: If you have any additional comments to share with the authors, please provide them here. Please also refer to our Reviewer’s guide on what makes a good review and pay specific attention to the different assessment criteria for the different paper categories: https://conferences.miccai.org/2025/en/REVIEWER-GUIDELINES.html
N/A
- Rate the paper on a scale of 1-6, 6 being the strongest (6-4: accept; 3-1: reject). Please use the entire range of the distribution. Spreading the score helps create a distribution for decision-making.
(3) Weak Reject — could be rejected, dependent on rebuttal
- Please justify your recommendation. What were the major factors that led you to your overall score for this paper?
This paper provide a Periapical Radiograph dataset. The novality of proposed model including the Multi-scale Wavelet Convolution Network (MWCN) and the Channel Fusion Attention is too weak. Neither dataset processing nor parameters are described.
- Reviewer confidence
Somewhat confident (2)
- [Post rebuttal] After reading the authors’ rebuttal, please state your final opinion of the paper.
Accept
- [Post rebuttal] Please justify your final decision from above.
The PRAD-10K dataset should be public at once.
Review #2
- Please describe the contribution of the paper
The paper introduces, PRAD-10K, the first and largest periapical radiography (PR) dataset, consisting of 10,000 images with corresponding expert annotations for nine distinct anatomical structures, lesions, and artificial restorations or medical devices. The dataset can be used for both PR segmentation and disease classification tasks. The authors also propose, PRNet, a PR image segmentation method, that utilizes novel Multi-scale Wavelent Convolution Network (MWCN) and Channel Fusion Attention (CFA) blocks within a U-Net architecture to address the multi-scale challenge in PR segmentation. Experimental results suggest that PRNet outperforms state-of-the-art medical image segmentation methods on the PRAD-10K dataset.
- Please list the major strengths of the paper: you should highlight a novel formulation, an original way to use data, demonstration of clinical feasibility, a novel application, a particularly strong evaluation, or anything else that is a strong aspect of this work. Please provide details, for instance, if a method is novel, explain what aspect is novel and why this is interesting.
- Novel dataset-PRAD10K: This is the first large PR dataset, consisting of 10,000 images with corresponding expert annotations.
- Novel PR segmentation method: The proposed PRNet innovatively combines both local and global features using both the MWCN modules in the U-Net encoder, and CFA modules in the skip connections. The MWCN modules consist of two convolution layers and two Wavelent convolution (WTConv) layers in parallel, with the conv. layers extracting local features and the WTConv layers extracting global features. The attention map in the CFA module is obtaining by averaging the input feature map at different scales in effect fusing both local and global features. The authors conducted ablation studies to validate each component.
- State-of-the-art performance on PRAD10K dataset: PRNet sets the SOTA performance on PRAD10K compared to typical and recently proposed SOTA medical image segmentation methods.
- Please list the major weaknesses of the paper. Please provide details: for instance, if you state that a formulation, way of using data, demonstration of clinical feasibility, or application is not novel, then you must provide specific references to prior work.
Dataset
- Limited details about the dataset: The data distribution is not mentioned: specifically, the number of patients, number of images per patient, number of images per annotated structure, number of disease cases, number of machines used to collected the dataset, and experience (#. of years) of the annotators.
- Limited diversity of the dataset: The dataset was collected from a single hospital and may contain inherent biases.
- Missing annotator variability assessment: The dataset was divided into two and each annotator processed one group, thus it is not clear how reliable the annotations are. While the authors mention that the annotators reviewed each other’s labels, this is not sufficient as it does not capture the less subtle differences in annotator variability. An assessment on both inter and intra-rator variability (even on just a fraction of the dataset) would be great.
- No benchmarking for the classification task: In Sect. 2 and Table 1, the authors mention that the dataset contains classification labels for periodontis, apical periodontis, and inadequate root canal fillings and can be used for classification task, however, this was not benchmarked in the experiments.
- Missing existing public PR: While the authors mentioned various existing publicly available datasets, the following very related PR dataset was not mentioned: Thalji et al. “Segmented X-ray image data for diagnosing dental periapical diseases using deep learning.” Data Brief. 2024 May
Proposed method:
- High performance variability across different structures: The authors propose PRNet to solve the multi-scale challenge, however, PRNet still performs differently across different structures, performing better on large structures such as bone and tooth, and poorly on implants and denture crowns. The paper does not provide an explanation for this performance difference.
- No comparison with existing periapical radiography segmentation methods: The following methods have been proposed for PR segmentation: 1) Ari et al. “Automatic Feature Segmentation in Dental Periapical Radiographs”. Diagnostics, 12(12), 3081. 2) Khan et al. “Automated feature detection in dental periapical radiographs by using deep learning. Oral Surgery, Oral Medicine, Oral Pathology and Oral Radiology, 131(6), 711–720. 3) Fatima et al., Deep Learning-Based Multiclass Instance Segmentation for Dental Lesion Detection. Healthcare 2023, 11, 347.
General
- No discussion on limitations of both the dataset and proposed method: As mentioned above, some limitations on the dataset include no quality assessment of the annotations, and lack of diversity in the dataset.
- Please rate the clarity and organization of this paper
Good
- Please comment on the reproducibility of the paper. Please be aware that providing code and data is a plus, but not a requirement for acceptance.
The authors claimed to release the source code and/or dataset upon acceptance of the submission.
- Optional: If you have any additional comments to share with the authors, please provide them here. Please also refer to our Reviewer’s guide on what makes a good review and pay specific attention to the different assessment criteria for the different paper categories: https://conferences.miccai.org/2025/en/REVIEWER-GUIDELINES.html
Major comments
- Missing relevant literature: While the introduction gives a comprehensive overview of segmentation methods in dentistry, it only mentions PR methods for bone loss and root fractures. A mention of the methods highlighted in the weakness section would provide a more complete and fair literature overview.
- Missing information about the dataset: The authors mention that they left out private information such that it is not included for research, it is very important to know where the dataset was collected for transparency and bias considerations.
- Clarification on MWCN module: How are the Global-local Feature Weighting Matrices (GFWM) selected? Are they trainable? Also, according to equation 1, the global feature fusion occurs once after the input features passing throught both convolution layers with two kernel sizes, but Fig. 3 shows that the fusion happens at both levels of convolution layers, once after k=5 and once after k=3.
- Implementation of SOTA methods: It is not clear to how the baseline methods were trained? Were they trained using the same setup mentioned in Sect. 4.1 or using the setup in their respective papers. This is important for fair comparison.
Minor comments
- Unclear conclusion of first paragraph in introduction: I did not understand the transition from challenges of PAN and CBCT, to, how DL techniques can enhance CAD in endodontics. A more chronological flow would be: highlight the different imaging modalities used, followed by limitations of PAN and CBCT, how PR addressess those limitations, and finally, highlight the challenges of PR that DL can solve.
- It would be great to have a reference supporting the statement that PR is the most frequently employed imaging technique in endodontics.
- Define WTConv. (Sect. 3.2)
- Rate the paper on a scale of 1-6, 6 being the strongest (6-4: accept; 3-1: reject). Please use the entire range of the distribution. Spreading the score helps create a distribution for decision-making.
(4) Weak Accept — could be accepted, dependent on rebuttal
- Please justify your recommendation. What were the major factors that led you to your overall score for this paper?
The paper proposes a novel large-scale dataset and a novel method segmentation method that achieves state-of-the-art performance on the dataset.
- Reviewer confidence
Very confident (4)
- [Post rebuttal] After reading the authors’ rebuttal, please state your final opinion of the paper.
Accept
- [Post rebuttal] Please justify your final decision from above.
The paper introduces a large-scale dataset and the authors have addressed my comments and minimal revision of the manuscript is required.
Review #3
- Please describe the contribution of the paper
The paper introduces a new dental dataset of periapical radiographs, while also including a number of benchmark methods and comparison partners to assess performance in various tasks.
- Please list the major strengths of the paper: you should highlight a novel formulation, an original way to use data, demonstration of clinical feasibility, a novel application, a particularly strong evaluation, or anything else that is a strong aspect of this work. Please provide details, for instance, if a method is novel, explain what aspect is novel and why this is interesting.
The dataset is new and comprehensive, well-annotated, and will be publicly available. Thus, the paper contributes to the larger MICCAI machine-learning ecosystem.
- Please list the major weaknesses of the paper. Please provide details: for instance, if you state that a formulation, way of using data, demonstration of clinical feasibility, or application is not novel, then you must provide specific references to prior work.
The paper does not suffer from any major shortcomings.
- Please rate the clarity and organization of this paper
Good
- Please comment on the reproducibility of the paper. Please be aware that providing code and data is a plus, but not a requirement for acceptance.
The authors claimed to release the source code and/or dataset upon acceptance of the submission.
- Optional: If you have any additional comments to share with the authors, please provide them here. Please also refer to our Reviewer’s guide on what makes a good review and pay specific attention to the different assessment criteria for the different paper categories: https://conferences.miccai.org/2025/en/REVIEWER-GUIDELINES.html
N/A
- Rate the paper on a scale of 1-6, 6 being the strongest (6-4: accept; 3-1: reject). Please use the entire range of the distribution. Spreading the score helps create a distribution for decision-making.
(5) Accept — should be accepted, independent of rebuttal
- Please justify your recommendation. What were the major factors that led you to your overall score for this paper?
Healthcare applications in machine learning can only thrive based on good, public datasets. The manuscript at hand makes a suitable foray in that direction and I am convinced that the dataset will prove highly relevant to the community.
- Reviewer confidence
Confident but not absolutely certain (3)
- [Post rebuttal] After reading the authors’ rebuttal, please state your final opinion of the paper.
Accept
- [Post rebuttal] Please justify your final decision from above.
All questions have been answered.
Author Feedback
On behalf of all co-authors of this paper, I would like to express our gratitude to the chairs and the three reviewers for their review work. Below, we address the reviewers’ comments:
To Reviewer 1: Thank you for taking the time to review our manuscript.
To Reviewer 2: Our paper has two main contributions: (1) we introduce PRAD-10K, the first and largest PR analysis dataset. Regarding the dataset processing methods, we have provided the specific collection and annotation process in Section 2. Preprocessing conducted before training has been described in Section 4.1. Regarding the detailed parameters of the dataset (also addressing Reviewer 3), We will add detailed descriptions in Section 2 or provide them on the github. (2) We propose PRNet, which consists of two key blocks: MWCN and CFA. In MWCN, we innovatively apply WTConv to medical tasks for its large receptive field and design a dual-size parallel convolution-WTConv hybrid structure to capture both local and global features. In the CFA, we introduced a multi-scale channel fusion mechanism and obtained feature maps containing multi-scale information for decoding through attention. The experimental results in Table 2 demonstrate that PRNet achieves leading performance in PR image segmentation tasks. The ablation study in Table 3 proves that each component of PRNet plays a crucial role in enhancing segmentation performance.
To Reviewer 3: Response to the dataset-related comments: For dataset distribution details, please refer to our response of the same question to Reviewer 2. Location information has been redacted for blind review and will be publicly available in the future. The expansion of the dataset’s multicenter diversity will be achieved in future work. To address the issue of inter-annotator variability, we have already contacted two chief physicians with 10 years of clinical experience from another authoritative hospital to initiate a new round of data annotation review. We will make every effort to ensure the accuracy of annotations before release the dataset. Due to space limitations, we focused on the more challenging segmentation task instead of classification. Intuitively, if the network can accurately segment categories such as apical periodontitis and root canal fillings at the pixel level, it should also perform well on the less complex classification tasks. We may propose separate tests for classification tasks in future work. Finally, we will update Table 1 to include the related dataset you mentioned.
Response to the method-related comments: The relatively poor performance of PRNet (and other models) on implants and dental crowns is likely due to feature confusion, as they often exhibit similar imaging characteristics. Additionally, large dental restorations can also resemble crowns, increasing misclassification risk. Addressing this issue is a key direction for future work, and we will include this analysis in Section 4.2. The three studies you mentioned primarily focus on clinical feasibility validation, employing UNet-based architectures for segmentation tasks that differ from ours. Notably, they lack publicly available implementations. Thus, experimental comparison may not meaningful, but we will cite them in the introduction to enrich the background of our work.
Response to other comments: GFWM are trainable and initialized to all-ones matrixes. Figure 3 shows the correct structure, while Eq.1 contains a typographical error; both will be corrected in the revised version. We used the same method mentioned in section 4.1 to conduct the training of all model, ensuring a fair comparison. We will include this description and modify the introduction section and add relevant references mentioned above in the revised version. We may not be able to include the definition of WTConv in the paper due to the page limitation. However, we have cited the relevant papers to ensure that readers can find the definition in the original publications.
Meta-Review
Meta-review #1
- Your recommendation
Invite for Rebuttal
- If your recommendation is “Provisional Reject”, then summarize the factors that went into this decision. In case you deviate from the reviewers’ recommendations, explain in detail the reasons why. You do not need to provide a justification for a recommendation of “Provisional Accept” or “Invite for Rebuttal”.
N/A
- After you have reviewed the rebuttal and updated reviews, please provide your recommendation based on all reviews and the authors’ rebuttal.
Accept
- Please justify your recommendation. You may optionally write justifications for ‘accepts’, but are expected to write a justification for ‘rejects’
N/A
Meta-review #2
- After you have reviewed the rebuttal and updated reviews, please provide your recommendation based on all reviews and the authors’ rebuttal.
Accept
- Please justify your recommendation. You may optionally write justifications for ‘accepts’, but are expected to write a justification for ‘rejects’
N/A
Meta-review #3
- After you have reviewed the rebuttal and updated reviews, please provide your recommendation based on all reviews and the authors’ rebuttal.
Accept
- Please justify your recommendation. You may optionally write justifications for ‘accepts’, but are expected to write a justification for ‘rejects’
Based on the rebuttal, I recoomend to accept this paper. But the link of the dataset should be provided in the camera ready version.