PMCanalSeg: A dataset for automatic segmentation of the pterygopalatine and mandibular canals from 3D CBCT images

Abstract

In orthognathic surgery, accurate segmentation of the pterygopalatine and mandibular canals in maxillofacial cone beam computed tomography (CBCT) scans is crucial. It provides critical information to prevent nerve damage during surgery and significantly reduces the risk of surgical complications. However, the high cost of data collection, strict patient privacy protection, and ethical constraints have hampered the performance of existing deep learning methods for pterygopalatine and mandibular canals segmentation, limiting their practical applicability in clinical settings. To address this challenge and advance the development of pterygopalatine and mandibular canal segmentation techniques in maxillofacial CBCT scans, we carefully constructed and made publicly available a large dataset for pterygopalatine and mandibular canal segmentation in maxillofacial CBCT scans. This dataset includes 191 patient cases and comprehensively covers the key anatomical structures of the maxillary pterygopalatine canal and the mandibular canal, both of which are crucial in orthognathic surgery. Notably, this dataset is the first to include data on the maxillary pterygopalatine canal, filling a significant gap in this field. The release of this dataset will greatly accelerate the development of deep learning-based segmentation methods, provide clinicians with more accurate reconstruction tools, and ultimately improve the safety and efficiency of surgical procedures.

Background & Summary

Dentomaxillofacial deformities refer to structural and functional abnormalities in the maxillofacial region, often resulting from uncoordinated skeletal development or abnormal tooth alignment¹. These deformities typically involve abnormalities in the teeth, jaws, and associated soft tissues, significantly impacting the patient’s physical and mental well-being. Patients with such conditions often require orthognathic surgery to improve occlusal function and restore facial aesthetics, which generally involves osteotomy of the maxilla and mandible². A critical aspect of osteotomy is the precise avoidance of the pterygopalatine canal and mandibular canal to minimize the risk of intraoperative bleeding and reduce the likelihood of postoperative complications^3,4. To enhance the safety and efficacy of orthognathic surgery, accurately segmenting complex anatomical structures like the pterygopalatine canal in the maxilla and the mandibular canal in the mandible using artificial intelligence technology has become a key research focus.

The traditional method for segmenting the pterygopalatine and mandibular canals is for doctors to use computer-aided medical imaging technology⁵ to manually mark the specific positions, directions and spatial relationships of nerves and blood vessels on each layer of CBCT images. However, this method is time-consuming and labor-intensive, heavily dependent on the doctor’s experience and technical level, and is easily affected by human errors and subjective judgment, which affects the accuracy and reliability of the segmentation results to a certain extent⁶.

In recent years, artificial intelligence (AI), particularly deep learning technology, has made significant advancements in the field of medical image segmentation^7–9. Through end-to-end learning, deep learning models can automatically extract feature representations from medical images and accurately identify segmentation boundaries, thus greatly enhancing the accuracy and efficiency of segmentation. Notable models such as U-Net¹⁰, UNETR¹¹, SwinUNETR¹², SegResNet¹³, and DynUNet¹⁴ have demonstrated outstanding performance in segmenting 2D and 3D medical images, including X-rays, CBCT scans, and MRIs. While these models have shown great success in medical image segmentation, their application to specialized tasks, such as the segmentation of the pterygopalatine canal and mandibular canal, typically requires high-quality annotated datasets for training and fine-tuning¹⁵. Therefore, the creation of high-quality, open-source datasets is essential, not only for training and optimizing segmentation models but also for providing a standardized benchmark that facilitates comparison and further improvement of segmentation algorithms across the research community.

Currently, there is a severe lack of public datasets and evaluation benchmarks for specific medical challenges, such as accurate segmentation of the pterygopalatine and mandibular canals in dento-maxillofacial CBCT scans. The shortage of data resources not only hinders the in-depth development and application of deep learning models in this specialized field, but also poses challenges for the objective evaluation of model performance and technical progress¹⁶. After a thorough review of the literature on mandibular canal segmentation, we found a large number of studies on this topic (as shown in Table 1); however, only four datasets are publicly available. Each subsequent version serves as an extension and improvement upon previous releases. The first dataset, Maxillo 2022, was created by the University of Modena and Reggio Emilia and contains 347 mandibular CBCT images, including 91 densely annotated images and 256 sparsely annotated images. The second dataset, ToothFairy 2023, which is an extension and improvement of this Maxillo dataset, contains 443 CBCT images of the mandible, including 153 densely annotated and 290 sparsely annotated images. This dataset specifically targets mandibular nerve segmentation tasks.

Table 1.

Detailed information of the related datasets.

Dataset	Year	Modality	Number of Image	Open Source
Al-Azhar University27	2012	CBCT	141	No
Korea University28	2018	CBCT	138	No
The Maxillo dataset29	2022	CBCT	347	Yes
The ToothFairy dataset30	2023	CBCT	443	Yes
The ToothFairy dataset v231	2024	CBCT	480 (50 for testing)	Yes
The ToothFairy dataset v323	2025	CBCT	532 (50 for testing)	Yes

The third dataset, ToothFairy 2024, further extends and refines the earlier released Maxillo and ToothFairy datasets. This updated version encompasses over 450 CBCT volumes for training and 50 volumes for testing, each fully annotated across 42 anatomical categories, including the background, jaw, inferior alveolar canal, and maxillary sinus, among others.

The latest release, ToothFairy 2025, supplements the 2024 edition by incorporating 52 additional CBCT scans collected using diverse scanners. These new volumes introduce 35 extra anatomical labels, notably including the pulp cavities of all 32 teeth, as well as the left and right incisive canals and the lingual canals. Although the ToothFairy dataset series now encompasses a total of 77 anatomical categories, including the mandibular nerve canal, it has consistently lacked annotations for the maxillary pterygopalatine canal. This drawback limits the broader applicability of the dataset series in comprehensive maxillofacial segmentation tasks and underscores the significance of our dataset.

To address these challenges, we created and released the first publicly available dataset, PMCanalSeg, specifically for segmenting the maxillary pterygopalatine canal and the mandibular canal. This dataset is used to train and evaluate the performance of deep learning segmentation models. It includes high-resolution 3D CBCT scans of 191 patients with voxel-level annotations provided by multiple medical experts. The dataset provides a diverse range of cases covering different age groups, genders, and various anatomical variations and pathological conditions. It is divided into two parts: one focuses on the anatomy of the pterygopalatine canal in the maxillary region, and the other focuses on the anatomy of the mandibular canal in the mandibular region. In addition, this paper introduces a series of experiments using advanced deep learning models to evaluate the accuracy and performance of the dataset. The experimental results show that our dataset achieves high segmentation accuracy for both the pterygopalatine and mandibular canals, providing clinicians with more accurate reconstruction tools, thereby improving the safety and efficiency of surgery.

Methods

The development of the PMCanalSeg dataset mainly includes four stages: Data collection, Data preprocessing, Data annotation, and Dataset organization. The main stages in the workflow are shown in Fig. 1. A detailed description of each stage in the workflow is as follows:

Fig. 1.

This diagram provides an overview of the dataset processing procedure, including Data collection, Data preprocessing, Data annotation, and Dataset organization.

Ethical Statement

This study was approved by the Ethics Committee of Jilin University (approval number: 2021038), with approval extending to the acquisition of imaging data, reconstruction of the patient’s oral 3D model, and the sharing and use of imaging data. To safeguard patient privacy, all personally identifiable information was carefully removed or anonymized in accordance with the General Data Protection Regulation (GDPR) and the Health Insurance Portability and Accountability Act (HIPAA). Specifically, anonymization included assigning randomized IDs to patient folders and removing sensitive DICOM tags that might lead to patient identification. All patients provided informed consent prior to data collection and explicitly agreed to the anonymization process and public data sharing.

Given the data collection period between 2021 and 2023, the ethical approval was confirmed to remain valid or was renewed appropriately. This study adheres strictly to medical ethics principles outlined in the Declaration of Helsinki and follows guidelines set by the Council for International Organizations of Medical Sciences (CIOMS). Additionally, robust safeguards have been implemented to prevent data misuse. These include strict controls on data access, limiting its usage exclusively to legitimate scientific research purposes. Moreover, to clearly define data-sharing permissions and restrictions, the dataset is licensed under a Creative Commons Attribution-NonCommercial-ShareAlike (CC-BY-NC-SA) license.

Dataset Overview

PMCanalSeg dataset is based on data collected by the Department of Max-illofacial Surgery, Jilin University Hospital of Stomatology, Changchun, China. It contains high-quality CBCT images of the head region captured by CBCT machines. CBCT machines are widely used in orthognathic surgery and imaging of maxillofacial anatomical structures due to their high imaging resolution, low radiation dose, and short scanning time^17–19. The focus area of the dataset is the maxilla and mandible, and the specific anatomical structures are the pterygopalatine canal and mandibular canal. The dataset contains CBCT scans of 191 patients who underwent or were scheduled for orthognathic surgery. All 191 cases are densely annotated, and the pterygopalatine canal and mandibular canal have accurate segmentation labels. To facilitate the training and evaluation of AI-based segmentation models, we divide the dataset into training, validation, and test sets in a ratio of 7:1:2.

Data collection

Given the potential risks associated with radiographic examinations, this study did not adopt a prospective trial; that is, no volunteers were recruited specifically for unnecessary examinations. Instead, we utilized the existing patient database resources of the hospital. Specifically, we collected data from patients aged between 17 and 35 years who visited the dental hospital between January 2021 and July 2023. The head data of these patients were obtained using cone beam computed tomography (CBCT) with a Neusoft Medical Systems scanner. The scanning parameters are detailed in Table 2. During the scans, patients were positioned supinely, with the upper and lower tooth apexes interposed and the midface axis aligned with the longitudinal axis of the CBCT scanner. Each patient’s dataset contained 300 to 400 .dcm (DICOM) files. To protect patient privacy, all identifiable personal information was anonymized, leaving only basic details such as gender, age, and scanning year accessible. Additionally, informed consent was obtained from all patients, explicitly agreeing to share their anonymized data for research and publication purposes. The anonymized patient information is documented in the information.xlsx file. Figure 2 illustrates an example of the three-dimensional head data obtained from the patients’ scans.

Table 2.

A summary of the scan parameters used.

Scan parameters	data
Reconstruction Interval	0.8mm
Tube Voltage	140KV
Tube Current	70mA
Slice Thickness	0.3125mm

Fig. 2.

From left to right, these are the cross-sectional, coronal, and sagittal images of a complete head scan CBCT.

Data preprocessing

To facilitate user access and support subsequent deep learning applications, the original DICOM format CBCT data was converted to NIfTI format. Data preprocessing was performed using SimpleITK²⁰ and 3D Slicer²¹. Specifically, the original DICOM files were imported into the 3D Slicer annotation software to reconstruct a 3D hard tissue model of the skull. The CT intensity threshold for bone reconstruction was set between 450 and 3071 Hounsfield units to generate an independent 3D skull model. To remove irrelevant structures, the spine was manually excluded using the cropping tool in 3D Slicer. The generated bone mask was then element-wise multiplied with the original CBCT volume to suppress non-bone structures such as soft tissue and cervical vertebrae. This process ensured that only relevant craniofacial bone components were retained.

Next, the 3D skull volume is divided into maxillary and mandibular regions based on the manually annotated labels of the maxillary and mandibular bones generated by SimpleITK. All preprocessed volumes and their corresponding label maps are saved in .nii.gz format and used as training data for subsequent deep learning tasks. The visualization of the final 3D processing result is shown in Fig. 3.

Fig. 3.

From left to right, they are the image of original 3D head, the image of preprocessed maxillary, and the image of preprocessed xiamandibular.

Data annotation

To obtain high-quality annotations, a team of four dental experts with extensive experience in maxillofacial surgery manually annotated each CBCT scan, producing dense voxel-level annotations for the pterygopalatine and mandibular canals. The annotation process involved several steps. First, two surgeons annotated each patient’s CBCT scan layer by layer, carefully delineating the boundaries of the pterygopalatine and mandibular canals and creating a label file suitable for model training. Once the preliminary annotations were completed, two additional surgeons conducted a thorough layer-by-layer verification, cross-checking the annotated positions, morphologies, and boundaries of the canals with the original scan data. This verification process ensured that the annotations were accurate and reliable for use in computer-assisted clinical trials. To quantitatively assess annotation reliability, we conducted an inter-observer consistency analysis on a randomly selected subset of 10 cases. The Dice similarity coefficient (DSC) was computed for both the pterygopalatine and mandibular canals, yielding an average DSC of 0.96  ± 0.02 and 0.98  ± 0.01, respectively. These results demonstrate a high degree of consistency among annotators, reflecting the reliability of the annotation process. The meticulous multi-stage workflow ensured accurate and standardized voxel-level labels, providing a solid foundation for downstream analyses, model training, and clinical applications. An example of the annotated pterygopalatine and mandibular canals is presented in Fig. 4.

Fig. 4.

The first row shows the annotated images of the pterygopalatine canal in the maxilla. From left to right, they are sample slices of the transverse section, coronal plane, and sagittal plane. The rightmost shows the annotated 3D pterygopalatine canal image. The second row shows the annotated images of the mandibular canal in the mandible. From left to right, they are sample slices of the transverse section, coronal plane, and sagittal plane. The rightmost shows the annotated 3D pterygopalatine canal image.

Data organization

After data annotation, multiple annotated 2D axial slices were combined to generate 3D data, which was then saved in .nii.gz format, ensuring the creation of accurate and comprehensive binary masks for each sample image. The samples and annotation data from 191 patients were organized into separate folders for each patient, with real IDs replaced by random numbers to protect privacy. Within each patient folder, “upper” and “lowe” subfolders were created to classify the corresponding samples and annotated data. The entire dataset is structured for ease of use, allowing readers to access it without the need for any additional conversion or modification.

Data Records

The PMCanalSeg dataset has been uploaded to Harvard Dataverse in a folder format and is available for download²². As shown in Fig. 5, the dataset is hierarchical. Each folder is named with the patient’s ID followed by an underscore and a number representing the patient’s visit number. Within each folder contains three subfolders: the upper folder contains maxillary samples and pterygopalatine canal labels, the lower folder contains mandibular samples and corresponding mandibular canal labels, and the skull folder contains the most original complete skull samples. In addition, the information.xlsx file provides the patient’s age, gender, and data acquisition time for each CBCT scan.

Fig. 5.

Structure of the data in the PMCanalSeg dataset.

Technical Validation

To demonstrate that our PMCanalSeg dataset can be effectively used for high-quality research and serve as a benchmark for future studies, we evaluated it using five representative AI-based segmentation models: U-Net¹⁰, UNETR¹¹, SwinUNETR¹², SegResNet¹³, and DynUNet¹⁴. A comprehensive assessment of both quantitative and qualitative performance was conducted using five evaluation metrics: IoU (Intersection over Union), DSC (DSC Similarity Coefficient), HD95 (95% Hausdorff Distance), ASD (Average Surface Distance), and RVE (Relative Volume Error).

To ensure statistical reliability and mitigate potential bias introduced by a single training-testing split, each model was trained and evaluated multiple times. Results are reported as mean  ± standard deviation to more fully reflect the consistency and robustness of model performance. All source code and training configurations are publicly available in our GitHub repository, allowing other researchers to replicate our experiments and further validate the quality of the PMCanalSeg dataset.

Quantitative Results on PMCanalSeg

Using the CBCT scan data of the maxilla and mandible, various models can be applied to directly segment the pterygopalatine and mandibular canals. The segmentation results for the pterygopalatine canal and mandibular canal are presented in Table 3 and Table 4, respectively.

Table 3.

Quantitative comparison of different methods for pterygopalatine canal segmentation in CBCT scans on the test set. footnote The best result is shown in bold and the second best result is underlined.

Methods	IoU↑	DSC↑	HD95(mm)↓	ASD(mm)↓	RVE↓
U-Net10	0.5989 ± 0.045	0.7412 ± 0.035	112.42 ± 12.30	8.72 ± 1.70	0.1650 ± 0.060
UNETR11	0.5684 ± 0.052	0.7151 ± 0.042	118.64 ± 11.70	9.68 ± 1.90	0.1954 ± 0.070
SwinUNETR12	$\underset{\_}{0.7127}\pm \underset{\_}{0.031}$	$\underset{\_}{0.8284}\pm \underset{\_}{0.025}$	$\underset{\_}{22.39}\pm \underset{\_}{6.40}$	$\underset{\_}{2.01}\pm \underset{\_}{0.60}$	$\underset{\_}{0.1136}\pm \underset{\_}{0.040}$
SegResNet13	0.6333 ± 0.039	0.7692 ± 0.030	23.10 ± 7.80	2.13 ± 0.70	0.1785 ± 0.050
DynUNet14	0.7167 ± 0.029	0.8310 ± 0.023	17.32 ± 5.90	1.17 ± 0.50	0.1106 ± 0.030

Table 4.

Quantitative comparison of different methods for mandibular canal segmentation in CBCT scans on the test set. footnoteThe best result is shown in bold and the second best result is underlined.

Methods	IoU↑	DSC↑	HD95(mm)↓	ASD(mm)↓	RVE↓
U-Net10	0.6994 ± 0.028	0.8219 ± 0.023	1.79 ± 0.91	0.68 ± 0.42	0.1374 ± 0.050
UNETR11	0.7463 ± 0.035	0.8492 ± 0.027	2.24 ± 1.18	$\underset{\_}{0.47}\pm 0.35$	0.2094 ± 0.060
SwinUNETR12	0.8465 ± 0.021	0.9243 ± 0.017	1.24 ± 0.68	0.32 ± 0.27	0.0728 ± 0.030
SegResNet13	0.6757 ± 0.032	0.8040 ± 0.025	1.92 ± 1.14	0.96 ± 0.52	0.2226 ± 0.070
DynUNet14	$\underset{\_}{0.8127}\pm \underset{\_}{0.025}$	$\underset{\_}{0.8955}\pm \underset{\_}{0.019}$	$\underset{\_}{1.48}\pm \underset{\_}{0.84}$	$\underset{\_}{0.55}\pm \underset{\_}{0.31}$	$\underset{\_}{0.0863}\pm \underset{\_}{0.040}$

According to the results of the pterygopalatine canal segmentation task given in Table 3, the DynUNet model performs well, with an IoU of 0.7167 and a DSC coefficient of 0.8310. At the same time, its boundary accuracy indicators HD95, ASD, and volume error RVE are all the lowest values, showing that its segmentation results are highly consistent with the true anatomical boundaries. However, compared with the mandibular canal segmentation task, the overall segmentation accuracy of the pterygopalatine canal is lower, with the highest IoU and DSC values of only 0.71 and 0.83, and the HD95 index is significantly higher, and the boundary segmentation effect is poor. This may be attributed to several interrelated factors.

First, the maxilla contains multiple cavities and foramina, which may interfere with model predictions by introducing structural ambiguity in the surrounding region. Second, the morphological structure of the pterygopalatine canal is inherently complex and exhibits significant inter-individual variation, lacking the consistent anatomical patterns observed in the mandibular canal. Third, due to its small diameter, poor visibility, and frequent overlap with adjacent structures, manual annotation of the pterygopalatine canal is more challenging, resulting in higher inter-observer variability and potential noise in the training data. Furthermore, the CBCT image quality in the region surrounding the pterygopalatine canal is often compromised by low soft-tissue contrast and beam hardening artifacts caused by adjacent air-filled sinuses, which can obscure canal boundaries and reduce model perception. Lastly, current general-purpose segmentation models—including DynUNet—may lack sufficient capacity to accurately learn and represent fine-grained, highly variable anatomical structures, particularly in the absence of dedicated geometric priors or multi-scale contextual modeling mechanisms. These factors have led to the unsatisfactory performance of the existing models in pterygopalatine canal segmentation, which needs further optimization.

As shown in Table 4, the SwinUNETR model performs well in the mandibular canal segmentation task, achieving the highest IoU (0.84645) and DSC (0.92428), as well as the smallest HD95 (1.23511mm), ASD (0.31611mm) and RVE (0.07283). These results show that the model performs better in mandibular canal segmentation than pterygopalatine canal segmentation, showing its superiority in segmenting the mandibular canal.

Although the DynUNet model performs best in the pterygopalatine canal segmentation task, in the mandibular canal segmentation task, the performance of DynUNet and UNETR is slightly lower than SwinUNETR, but still maintains high accuracy and strong spatial overlap, showing considerable competitiveness. In contrast, U-Net and SegResNet perform poorly on all five evaluation metrics, with low accuracy and large segmentation errors, indicating that their performance in both pterygopalatine canal segmentation and mandibular canal segmentation needs to be further improved. In future work, we plan to explore advanced data augmentation strategies and combine domain-specific architectural enhancements (such as multi-scale feature fusion or anatomical-based attention modules) to improve segmentation performance.

Qualitative Results on PMCanalSeg

We present visual comparisons between the predicted labels and ground truth labels. Figure 6 shows a 3D visualization of the pterygopalatine canal segmentation, while Fig. 7 illustrates a 3D visualization of the mandibular canal segmentation. These figures compare the ground truth labels with the predicted labels, highlighting the differences. The results demonstrate that while the models have achieved satisfactory segmentation for the mandibular canal, the segmentation performance for the pterygopalatine canal is slightly less inferior. This difference may be attributed to the inherent complexity of the pterygopalatine canal’s structure, which is more challenging to learn and predict compared to the mandibular canal.

Fig. 6.

3D image visualization of the pterygopalatine canal segmentation results. The ground truth label is on the far left. The first row is the model’s predictions (shown in blue), and the second row is an image of the predictions overlapped with the ground truth labels. The overlap of the labels has the following color codes: red for true positives, green for false positives, and yellow for false negatives. The third row is a magnified front view of the overlapping image, and the fourth row is a magnified side view of the overlapping image.

Fig. 7.

3D image visualization of mandibular canal segmentation results. The ground truth label is on the far left. The first row is the model’s predictions (shown in blue), and the second row is an image of the predictions overlapped with the ground truth labels. The overlap of the labels has the following color codes: red for true positives, green for false positives, and yellow for false negatives. The third row is a magnified front view of the overlapping image, and the fourth row is a magnified side view of the overlapping image.

Comparison between PMCanalSeg and other datasets

We conducted further experiments to compare PMCanalSeg with other existing datasets. Specifically, we selected the representative ToothFairy dataset for comparison, we employed three state-of-the-art segmentation models - nnU-Net, nnFormer, and TransU-Net - trained and evaluated separately on both PMCanalSeg and ToothFairy datasets. The experimental results, summarized in Table 5 and 6.

Table 5.

The results are obtained from the published ToothFairy dataset study²³, which quantitatively compared the performance of various mandibular canal segmentation methods on CBCT scans from the test set. footnote The best results are highlighted in bold, and the second-best results are indicated with underlines.

Methods	DSC↑	HD95(mm)↓
nnU-Net24	0.7134	29.11
nnFormer25	0.7228	10.06
TransU-Net26	0.8196	11.99

Table 6.

Quantitative comparison of different methods for mandibular canal segmentation in CBCT scans on the test set. footnote The best results are in bold and the second best are underlined.

Methods	Segmentation Manner	IoU↑	DSC↑	HD95(mm)↓	ASD(mm)↓	RVE↓
nnU-Net24	2D	0.84644	0.91645	1.05033	0.27156	0.07688
	3D-lowres	0.71314	0.83181	0.83181	0.50965	0.20567
	3D-fullres	0.86623	0.92818	1.0000	0.22373	0.08603
nnFormer25	3D	0.85188	0.91924	5.25220	3.73909	0.04916
TransU-Net26	3D	0.72317	0.83759	2.07309	0.53256	0.16459

The mandibular canal segmentation performance was evaluated on the respective test sets of the two datasets. For the ToothFairy dataset, the published results reported in²³ are summarized in Table 5, where nnU-Net achieved a DSC of 0.7134 and an HD95 of 29.11 mm; nnFormer achieved a DSC of 0.7228 and an HD95 of 10.06 mm; and TransU-Net achieved a DSC of 0.8196 and an HD95 of 11.99 mm. For the PMCanalSeg dataset, the experimental results are presented in Table 6. The nnU-Net model yielded DSC values of 0.91645, 0.83181, and 0.92818 for the 2D, 3D-lowres, and 3D-fullres configurations, respectively, with corresponding HD95 values of 1.05033 mm, 0.83181 mm, and 1.0000 mm. nnFormer obtained a DSC of 0.91924 and an HD95 of 5.25220 mm, while TransU-Net achieved a DSC of 0.83759 and an HD95 of 2.07309 mm.

Experimental results show that the models achieve higher performance metrics on the PMCanalSeg dataset compared with their results on the ToothFairy dataset in mandibular canal segmentation. However this does not necessarily imply an inherent advantage of PMCanalSeg in mandibular canal segmentation, since the two datasets differ in several aspects, which may collectively affect the model’s training results. First, PMCanalSeg employs a fine voxel-level annotation process with multiple rounds of iterative verification by several clinical experts, which may reduce label noise and make the mandibular canal boundaries more consistent. Second, this dataset introduces a multi-stage quality control mechanism during annotation, including automatic consistency checks and layer-by-layer review by senior clinicians, which may reduce annotation discrepancies and thus improve the stability of model learning. Finally, PMCanalSeg provides pre-separated maxillary and mandibular volumes, allowing the model to learn within a more focused structural range, unaffected by irrelevant anatomical regions; this organizational approach may also influence the final performance.

Furthermore, inherent differences in imaging orientation, spatial resolution, and scanning protocols between the two datasets may also lead to performance inconsistencies between the two datasets. After visually comparing ToothFairy and PMCanalSeg in 3D Slicer, we found significant domain differences between the two in spatial coordinate systems, acquisition methods, and anatomical annotation standards. Different scanning devices, voxel spacing, patient positioning, and imaging parameters can all cause inconsistencies in spatial context, hindering model generalization across datasets. This phenomenon is common in multicenter medical imaging studies, reflecting differences in dataset distribution rather than model capability or annotation quality. Therefore, these underlying factors, rather than differences in dataset quality themselves, are likely the cause of inconsistent model performance across the two datasets.

For the reasons stated above, our goal is not to conclude that the PMCanalSeg dataset is superior to other datasets, but rather to emphasize that the observed performance differences stem from differences in data features, rather than differences in the quality of the datasets themselves. Therefore, our current contribution lies in establishing PMCanalSeg as a standardized benchmark for fairly evaluating algorithms within the dataset.

Beyond benchmarking objectives, PMCanalSeg brings new clinical and technical value to the field. Particularly noteworthy is its inclusion, for the first time, of annotated data for the maxillary pterygopalatine canal, filling a significant gap in the field. This dataset performs well across multiple evaluation metrics, supports the training and testing of various deep learning algorithms, enables robust validation and model enhancement, and helps clinicians accurately segment the pterygopalatine and mandibular canals. By publicly making this high-quality segmentation data available, our work aims to advance deep learning research in this field and lay a more solid foundation for future medical diagnosis and treatment.

Author contributions

G.L. and L.W. contributed to data curation, methodology development, software implementation, original draft preparation, and manuscript revision. G.W. provided data oversight, project supervision, and critical revisions. Y.L. and R.M. led manuscript refinement, funding acquisition, project oversight, and administrative coordination. All authors participated in manuscript review and approved the final version.

Data availability

The PMCanalSeg dataset is available at 10.7910/DVN/RTIGTP.

Code availability

All the code used for dataset loading, data analysis, and experiments in this article is found in the GitHub repository at https://github.com/lgh010319/PMCSeg. When analyzing the data, the evaluation metrics of the experiment should be followed, and we encourage sharing of code and models to promote reproducibility.

Competing interests

The authors declare no competing interests.