Background & Summary
Mandibular defects, a common form of craniomaxillofacial bone defect, can be classified using the classic HCL system, as shown in Fig. 1b and Table 1, which includes unilateral defects, bilateral defects, and complex defects crossing the midline1. The primary etiologies of mandibular defects include tumor resection[2](https://www.nature.com/articles/s41597-025-06048-8#ref-CR2 “Ch…
Background & Summary
Mandibular defects, a common form of craniomaxillofacial bone defect, can be classified using the classic HCL system, as shown in Fig. 1b and Table 1, which includes unilateral defects, bilateral defects, and complex defects crossing the midline1. The primary etiologies of mandibular defects include tumor resection[2](https://www.nature.com/articles/s41597-025-06048-8#ref-CR2 “Chatterjee, D. et al. Reconstruction of complex oro-mandibular defects by four different modifications of free fibula osteomyocutaneous flap: A prudent alternative to multiple flaps. J Plast Reconstr Aesthet Surg 75, 3346–3355, https://doi.org/10.1016/j.bjps.2022.04.060
(2022).“), severe trauma[3](https://www.nature.com/articles/s41597-025-06048-8#ref-CR3 “Frensilli, J. A., Kornblut, A. D. & Tenen, C. Reconstruction of a mandible after shotgun trauma: report of case. J Am Dent Assoc 110, 49–51, https://doi.org/10.14219/jada.archive.1985.0287
(1985).“), and osteomyelitis of the jaw[4](https://www.nature.com/articles/s41597-025-06048-8#ref-CR4 “Xie, Q., Jiang, X. & Huang, X. Distraction osteogenesis application in bone defect caused by osteomyelitis following mandibular fracture surgery: a case report and literature review. BMC Musculoskelet Disord 25, 813, https://doi.org/10.1186/s12891-024-07922-z
(2024).“). These defects not only significantly impair essential functions, such as mastication and speech, but also alter facial aesthetics, which may potentially impact patients’ social functioning and psychological well-being[5](#ref-CR5 “Coletti, D. P., Ord, R. & Liu, X. Mandibular reconstruction and second generation locking reconstruction plates: outcome of 110 patients. Int J Oral Maxillofac Surg 38, 960–963, https://doi.org/10.1016/j.ijom.2009.03.721
(2009).“),6,7. As a result, the reconstruction of mandibular defects plays a critical role in oral and maxillofacial surgery.
Fig. 1
CT data processing workflow. (a): Schematic diagram of the data processing procedure. (b): Classical HCL classification standard. (c): Anatomical landmarks localization and reference planes generation. (d): Smooth connection between the reconstructed part and the residual mandible.
Currently, 3D reconstruction visualization and virtual surgical planning (VSP) have become reliable techniques to achieve precise reconstruction using high-resolution computed tomography (CT) images[8](#ref-CR8 “Annino, D. J. Jr. et al. Accuracy and outcomes of virtual surgical planning and 3D-printed guides for osseous free flap reconstruction of mandibular osteoradionecrosis. Oral Oncol 135, 106239, https://doi.org/10.1016/j.oraloncology.2022.106239
(2022).“),[9](#ref-CR9 “Ettinger, K. S. et al. Patient-specific Implants Improve Volumetric Surgical Accuracy Compared to Stock Reconstruction Plates in Modern Paradigm Virtual Surgical Planning of Fibular Free Flaps for Head and Neck Reconstruction. J Oral Maxillofac Surg 82, 1311–1328, https://doi.org/10.1016/j.joms.2024.06.166
(2024).“),[10](#ref-CR10 “Hurley, C. M. et al. Current trends in craniofacial reconstruction. Surgeon 21, e118–e125, https://doi.org/10.1016/j.surge.2022.04.004
(2023).“),[11](https://www.nature.com/articles/s41597-025-06048-8#ref-CR11 “Lo, L. J. & Lin, H. H. Applications of three-dimensional imaging techniques in craniomaxillofacial surgery: A literature review. Biomed J 46, 100615, https://doi.org/10.1016/j.bj.2023.100615
(2023).“). Typically, clinical CT imaging data is first processed to perform 3D reconstruction through medical image processing technology. Subsequently, anatomical landmarks serve as guides to establish reference planes. Finally, a complete morphology is reconstructed based on the reference planes and mandibular curvature, and the design is finalized by integrating specific treatment strategies[12](https://www.nature.com/articles/s41597-025-06048-8#ref-CR12 “Tatti, M. et al. Segmental Mandibulectomy and Mandibular Reconstruction with Fibula-Free Flap Using a 3D Template. J Pers Med 14 https://doi.org/10.3390/jpm14050512
(2024).“),13. However, significant limitations still exist in the design workflow. On the one hand, the precise control of spatial positioning and angles within the model design interface remains challenging13. In order to achieve the required configuration, repeated adjustments and measurements are often necessary, which consumes significant time and effort and results in low efficiency. On the other hand, for complex cases, traditional mirroring techniques fail to reconstruct the missing volume located between the left and right sides when the defect involves the midsagittal line14. These cases require further manual corrections, such as adjustments to curvature, thickness, and smoothing, which significantly increase the time required[15](https://www.nature.com/articles/s41597-025-06048-8#ref-CR15 “Tantisatirapong, S. et al. The simplified tailor-made workflows for a 3D slicer-based craniofacial implant design. Sci Rep 13, 2850, https://doi.org/10.1038/s41598-023-30117-w
(2023).“). Moreover, these tasks are highly dependent on the designer’s skill and experience, creating a steep learning curve for surgeons.
In craniofacial complexes, machine learning, especially deep learning models, can extract underlying patterns from large datasets of CT images and X-rays[16](https://www.nature.com/articles/s41597-025-06048-8#ref-CR16 “Heo, M.-S. et al. Artificial intelligence in oral and maxillofacial radiology: what is currently possible? Dentomaxillofacial Radiology 50, 20200375, https://doi.org/10.1259/dmfr.20200375
(2021).“), and predict complete bone shapes from incomplete or defective input data, thereby assisting in diagnosis, planning, and automated implant design[17](#ref-CR17 “Li, J. et al. Synthetic skull bone defects for automatic patient-specific craniofacial implant design. Sci Data 8, 36, https://doi.org/10.1038/s41597-021-00806-0
(2021).“),[18](#ref-CR18 “Li, J. et al. Sparse convolutional neural network for high-resolution skull shape completion and shape super-resolution. Sci Rep 13, 20229, https://doi.org/10.1038/s41598-023-47437-6
(2023).“),[19](#ref-CR19 “Mazzocchetti, S. et al. Neural shape completion for personalized Maxillofacial surgery. Sci Rep 14, 19810, https://doi.org/10.1038/s41598-024-68084-5
(2024).“),[20](#ref-CR20 “Wodzinski, M. et al. Deep learning-based framework for automatic cranial defect reconstruction and implant modeling. Comput Methods Programs Biomed 226, 107173, https://doi.org/10.1016/j.cmpb.2022.107173
(2022).“),[21](#ref-CR21 “Wu, C. T., Yang, Y. H. & Chang, Y. Z. Three-dimensional deep learning to automatically generate cranial implant geometry. Sci Rep 12, 2683, https://doi.org/10.1038/s41598-022-06606-9
(2022).“),[22](#ref-CR22 “Chuxi, Z., Xinkang, Z., Xiaokun, D., Shilei, Z. & Xinrong, C. CMF defects database: A craniomaxillofacial defects dataset and a data-driven repair method. Biomedical Signal Processing and Control 91, 105939, https://doi.org/10.1016/j.bspc.2023.105939
(2024).“),[23](#ref-CR23 “Han, B. et al. Statistical and individual characteristics-based reconstruction for craniomaxillofacial surgery. International Journal of Computer Assisted Radiology and Surgery 17, 1155–1165, https://doi.org/10.1007/s11548-022-02626-y
(2022).“),[24](#ref-CR24 “Jie, B. et al. Automatic virtual reconstruction of maxillofacial bone defects assisted by ICP (iterative closest point) algorithm and normal people database. Clinical Oral Investigations 26, 2005–2014, https://doi.org/10.1007/s00784-021-04181-3
(2022).“),[25](#ref-CR25 “Li, J. et al. Automatic skull defect restoration and cranial implant generation for cranioplasty. Medical Image Analysis 73, 102171, https://doi.org/10.1016/j.media.2021.102171
(2021).“),[26](#ref-CR26 “Modabber, A. et al. Evaluation of a novel algorithm for automating virtual surgical planning in mandibular reconstruction using fibula flaps. J Craniomaxillofac Surg 47, 1378–1386, https://doi.org/10.1016/j.jcms.2019.06.013
(2019).“),[27](https://www.nature.com/articles/s41597-025-06048-8#ref-CR27 “Zhong, C., Xiong, Y., Tang, W. & Guo, J. A Stage-Wise Residual Attention Generation Adversarial Network for Mandibular Defect Repairing and Reconstruction. Int J Neural Syst 34, 2450033, https://doi.org/10.1142/s0129065724500333
(2024).“). In the reconstruction planning process based on deep learning, a large volume of clinical defect data derived from real cases has the potential to significantly enhance the adaptability of AI in handling complex scenarios. Moreover, the mandible exhibits a complex and irregular shape, with considerable morphological variability among patients[28](https://www.nature.com/articles/s41597-025-06048-8#ref-CR28 “Liew, S. et al. Consensus on Changing Trends, Attitudes, and Concepts of Asian Beauty. Aesthetic Plast Surg 40, 193–201, https://doi.org/10.1007/s00266-015-0562-0
(2016).“),[29](https://www.nature.com/articles/s41597-025-06048-8#ref-CR29 “Martone, A. L. Anatomy of facial expression and its prosthodontic significance. The Journal of Prosthetic Dentistry 12, 1020–1042, https://doi.org/10.1016/0022-3913(62)90158-0
(1962).“), which further complicates the development of autonomous reconstruction algorithms. Therefore, there is an urgent need for a dataset of clinical mandibular defects that can support the development of intelligent algorithms and address the current challenges.
In the field of craniofacial reconstruction, existing datasets[17](#ref-CR17 “Li, J. et al. Synthetic skull bone defects for automatic patient-specific craniofacial implant design. Sci Data 8, 36, https://doi.org/10.1038/s41597-021-00806-0
(2021).“),[18](#ref-CR18 “Li, J. et al. Sparse convolutional neural network for high-resolution skull shape completion and shape super-resolution. Sci Rep 13, 20229, https://doi.org/10.1038/s41598-023-47437-6
(2023).“),[19](#ref-CR19 “Mazzocchetti, S. et al. Neural shape completion for personalized Maxillofacial surgery. Sci Rep 14, 19810, https://doi.org/10.1038/s41598-024-68084-5
(2024).“),[20](#ref-CR20 “Wodzinski, M. et al. Deep learning-based framework for automatic cranial defect reconstruction and implant modeling. Comput Methods Programs Biomed 226, 107173, https://doi.org/10.1016/j.cmpb.2022.107173
(2022).“),[21](#ref-CR21 “Wu, C. T., Yang, Y. H. & Chang, Y. Z. Three-dimensional deep learning to automatically generate cranial implant geometry. Sci Rep 12, 2683, https://doi.org/10.1038/s41598-022-06606-9
(2022).“),[22](#ref-CR22 “Chuxi, Z., Xinkang, Z., Xiaokun, D., Shilei, Z. & Xinrong, C. CMF defects database: A craniomaxillofacial defects dataset and a data-driven repair method. Biomedical Signal Processing and Control 91, 105939, https://doi.org/10.1016/j.bspc.2023.105939
(2024).“),[23](#ref-CR23 “Han, B. et al. Statistical and individual characteristics-based reconstruction for craniomaxillofacial surgery. International Journal of Computer Assisted Radiology and Surgery 17, 1155–1165, https://doi.org/10.1007/s11548-022-02626-y
(2022).“),[24](#ref-CR24 “Jie, B. et al. Automatic virtual reconstruction of maxillofacial bone defects assisted by ICP (iterative closest point) algorithm and normal people database. Clinical Oral Investigations 26, 2005–2014, https://doi.org/10.1007/s00784-021-04181-3
(2022).“),[25](#ref-CR25 “Li, J. et al. Automatic skull defect restoration and cranial implant generation for cranioplasty. Medical Image Analysis 73, 102171, https://doi.org/10.1016/j.media.2021.102171
(2021).“),[26](#ref-CR26 “Modabber, A. et al. Evaluation of a novel algorithm for automating virtual surgical planning in mandibular reconstruction using fibula flaps. J Craniomaxillofac Surg 47, 1378–1386, https://doi.org/10.1016/j.jcms.2019.06.013
(2019).“),[27](https://www.nature.com/articles/s41597-025-06048-8#ref-CR27 “Zhong, C., Xiong, Y., Tang, W. & Guo, J. A Stage-Wise Residual Attention Generation Adversarial Network for Mandibular Defect Repairing and Reconstruction. Int J Neural Syst 34, 2450033, https://doi.org/10.1142/s0129065724500333
(2024).“) cover most of the neurocranium and splanchnocranium regions, such as the skull, midface and mandible, as shown in Table 2. However, datasets specifically focused on the mandible remain limited. Furthermore, most of these studies rely on artificially generated defects, with only a small portion of actual clinical defect data being included. Significant differences exist between clinically derived defects and artificially generated defects. Regardless of the underlying etiology, defect boundaries exhibit rough and irregular cross-sections, which is a prominent characteristic of mandibular defects in clinical practice. In contrast, artificially generated defect data, generated and rendered by computer programs, produce smooth and straight cross-sections, which is a simplified approximation of clinical defects. Additionally, in actual clinical scenarios, a normal condyle-glenoid fossa positional relationship is often absent in defective mandibles, significantly complicating the identification of the defect’s true morphology. Moreover, the process of artificially introducing defects may involve unintentional subjective biases, which can propagate into trained AI models and lead to uncertainties in real-world applications. Therefore, algorithms trained on artificially generated defects often fail to effectively respond to the complex and interfering factors present in real-world cases, thus limiting their applicability in clinical scenarios. To address this limitation, ground truth data derived from clinical practice is particularly crucial, yet a comprehensive, large-scale clinical defect dataset specifically focused on mandibular reconstruction has not been contributed by previous studies.
This manuscript presents the first version of the Mandibular Defect Dataset, which comprises real clinical defect models in STL and NRRD format. The CT scanners used to acquire the data are from the Ninth People’s Hospital, Shanghai Jiao Tong University School of Medicine, with all machines subjected to regular quality assurance checks. The models were generated following standardized workflows designed to meet clinical requirements, ensuring sufficient resolution and usability. Each original dataset underwent image quality assessment, and experienced oral and maxillofacial surgeons contributed diagnoses of defect classification and annotations for each model. Additionally, statistical analyses of etiologies have been included.
In this work, a clinically derived dataset of mandibular defects is established, which is the first relevant dataset to the best of our knowledge. The established dataset aims to develop AI models for virtual shape completion tailored to various types of mandibular defects, which will offer intelligent solutions for personalized mandibular reconstruction. Specifically, by uploading 3D models derived from clinical CT data of patients with mandibular defects, the trained deep learning framework can autonomously generate virtually reconstructed mandibular models. In addition to rapidly generating implantable structures for 3D printing, this model is expected to enhance the efficiency of designing personalized reconstruction plans. Taking the commonly used vascularized fibula osteomyocutaneous flap as an example, the virtual model provides critical guidance on the placement and angulation of fibula segments, effectively reducing time costs and potential errors associated with iterative manual adjustments.
Methods
Ethical approval and inclusion/exclusion criteria
This study was approved by the institutional review board and the ethics committees of Shanghai Ninth People’s Hospital, Shanghai Jiao Tong University School of Medicine (SH9H-2021-T66-3). Due to the retrospective nature of the study, the requirement for oral or written informed consent from patients was waived. All methods related to the study were conducted in accordance with the Declaration of Helsinki and all relevant guidelines and regulations. The study included 203 patients who received diagnosis or treatment at Shanghai Ninth People’s Hospital between January 2020 and October 2024. The inclusion criteria were patients diagnosed with “mandibular defect” within the hospital. The exclusion criteria were defined as patients unable to undergo maxillofacial CT scans for various reasons, such as financial constraints, geographic remoteness preventing follow-up, or psychological resistance.
Data collection, processing, and quality assessment
The CT data was generated using the Discovery CT750 HD scanner manufactured by GE Medical Systems according to standard protocols. The data were initially provided in DICOM format. During the quick check, data of very poor quality was excluded from the review, such as images blurred due to obvious movement of the lower jaw during scanning, or data with a large number of artifacts masking anatomical structures, making them difficult to interpret. 147 sets of CT data involving skull or maxillary defects were ultimately included and retained, which focuses specifically on mandibular defects.
The data processing workflow was shown in Fig. 1. DICOM data was uploaded into the design software ProPlan CMF 3.0 (Materialise, Leuven, Belgium), where experienced surgeons, from the Department of Oral and Cranio-Maxillofacial Surgery at Shanghai Ninth People’s Hospital, affiliated with Shanghai Jiao Tong University School of Medicine, performed operations, analysis, and annotations. Annotations and analyses were conducted by senior attending physicians with over five years of clinical experience, supported by residents and medical students proficient in craniofacial anatomy who participated in statistical reviews and data verification during the annotation process. Specifically, the “Segmentation” module was utilized to perform semi-automated 3D reconstruction within a predefined threshold range (226–3071 HU). The software’s built-in “Remove Parts” module was then used for manual layer-by-layer editing to separate the maxilla and mandible, identify the defective mandible, and conduct HCL classification diagnoses.
During this process, data cleaning was also performed. Artifacts, typically caused by specific conditions or objects, result in image distortion30. On the one hand, there are structures that are easy to identify and segment, such as CT tables, nasogastric tubes, endotracheal tubes, drainage tubes, or packaging materials. On the other hand, artifacts caused by metal dental restorations, orthodontic brackets, and titanium plates or screws implanted after tumor resection, which are often adjacent to the defect area, can directly affect the accuracy of boundary delineation and annotation quality.
All images were carefully examined layer by layer to ensure that the final exported models and defect annotations were minimally affected by artifacts. However, despite these efforts, such objective conditions may still have some impact on mandibular segmentation. Therefore, image quality was categorized based on statistical analysis, as shown in Fig. 2. Images without obvious artifact interference were classified as high-quality. Images containing only artifacts from dental restorations or metallic orthodontic brackets are categorized as medium-quality. Images with artifacts caused by titanium plates and screws were classified as low-quality.
Fig. 2
Data quality analysis and statistics.
Different quality categories of data can be used for distinct purposes. For example, high-quality data can be utilized to analyze the anatomical structure of non-healthy mandibles. Medium-quality data, which often places higher demands on the segmentation of the maxillary and mandibular dentition, can be used to train or optimize AI classifiers for automated mandibular segmentation. Although low-quality images containing rigid internal fixation structures interfere with defect boundary delineation in this dataset, they represent real-world clinical scenarios that closely align with treatment strategies. These images are worth further discussion in future work and hold potential for applications in clinical review, teaching demonstrations, and skill training.
Manual annotation
Statistical analysis was conducted on the segmented mandibular defect models, including the classification of defect extent in Fig. 3a, the presence of concurrent defects in other regions in Fig. 3b, and the necessity of managing the mandibular remnants before reconstruction in Fig. 3c, such as localized bone trimming or repositioning. Additionally, the etiology corresponding to each case of mandibular defect was documented, as shown in Fig. 3d, providing valuable insights for potential future clinical cohort studies.
Fig. 3
Mandibular Defect Dataset analysis. (a): Defect classification statistics. (b): Presence of other cranio-maxillofacial defects. (c): Pre-reconstruction management requirements. (d): Defect source statistics.
Before initiating the design process, cephalometric analysis based on anatomical landmarks will be used to assist in marking reference planes, where precise positioning of the facial midline is crucial, as shown in Fig. 1c.
Within the software, the condylar structures are rotated and repositioned to improve model symmetry as much as possible, placing the mandibular remnants in positions closest to their normal anatomical locations, as shown in Fig. 4. Two situations for repositioning were summarized. (1) When pre-defect imaging is available, it can be imported into the design interface in STL format, with the edge contours serving as the gold standard for realignment. (2) When diseases or trauma have impacted the mandibular position before patients seek medical attention, reference imaging data or occlusal relationships are often unavailable for most patients. Therefore, repositioning is based on the intercuspal position (ICP) of the remaining teeth and the relative position of the condyle-glenoid fossa in the sagittal and coronal planes, with mirror correction performed using the facial midline reference plane. It is worth noting that this step does not overemphasize achieving absolute aesthetic symmetry, as some patients exhibit physiological characteristics such as occlusal plane canting or asymmetry in mandibular angle size.
Fig. 4
Repositioning of mandibular segments. (a) Virtual mandible after repositioning: original model (gray, semi-transparent), mirrored reference model (yellow, semi-transparent), and mandibular remnants (green, right; red, left, both opaque). (b) Symmetry assessment during repositioning using the mirrored reference model.
In addition, mandibular models with localized bone destruction require trimming of irregular and rough bone edges prior to reconstruction. Due to differences in the severity of benign or malignant conditions among patients and variability in clinical experience among physicians, the amount of bone removal during manual design often shows significant variation. To minimize the impact of this factor on data quality, the annotation process did not consider specific clinical treatment requirements but strictly adhered to the following two principles: (1) Avoidance of critical structures, including tooth roots and the course of major blood vessels and nerve bundles. (2) Minimal bone removal, solely aimed at smoothing thin, irregular, or poor-quality bone at the defect edges without compromising the overall authenticity of the defect.
The surgeon then determines whether the defect is distributed unilaterally or bilaterally relative to the facial midline. For defects located unilaterally at the midline, the mirror tool provided by the software typically generates a reliable reference model. However, for defects spanning bilaterally across the midline, conventional mirroring techniques are not applicable. In such cases, reconstruction relies solely on clinical expertise and pre-defect CT images to manually complete the defect, as shown in Fig. 5. Specifically, the procedure is carried out using Geomagic Studio 2013 (64-bit) (Geomagic Inc., Morrisville, NC). Defect boundaries are extracted from the cutting ends of the preprocessed model to form constraint curves that define the reconstruction region. The Interactive Surface Patches tool and the Mesh Bridging tool in the software are applied stepwise to complete the missing volume. The positions and shapes of control points are adjusted manually while preserving the original edge trajectories and curvature, as far as possible. After the patches are created, local smoothing and mesh repair are performed to ensure continuity of normals and curvature between the patches and the residual bone segments to achieve a watertight model. Mesh repair includes hole filling, removal of self-intersections, and elimination of nonmanifold elements. Finally, seams and local deviations are inspected, and small regions with concentrated errors are refined by smoothing and remeshing. The postprocessed model, which represents the final virtual reconstructed mandibular model, is then exported, as shown in Fig. 1d.
Fig. 5
Manual reconstruction of midline-crossing defects using Geomagic Studio 2013 (64-bit). (a–c): The reference mirror model was refined through cropping, segmentation, and adjustment. (d): Manual completions for unreconstructed regions based on edge contours and clinical expertise. (e): Postprocessed virtual mandibular model after edge optimization.
Data anonymization
STL files generated in the preceding steps are converted to NRRD using 3D Slicer 5.8.0 (Harvard University, Boston, USA). All shared data are finally provided as STL files and corresponding fully de-identified volumetric images in compressed NRRD format. No DICOM files are shared. All protected health information and acquisition identifiers are removed. Volumes are cropped to the region of interest, so that no soft-tissue facial features are retained. The NRRD headers retain only spatial information required for analysis and include no dates, names, IDs, or site-specific strings. Case identifiers cannot be linked to clinical records. These measures minimize re-identification risk while preserving utility for mandibular reconstruction research.
Data Records
The dataset is available at https://doi.org/10.7303/SYN69090156[31](https://www.nature.com/articles/s41597-025-06048-8#ref-CR31 “Wu, J. et al. A Mandibular Defect Dataset for Autonomous Reconstruction Planning in Oral and Maxillofacial Surgery. Synapse https://doi.org/10.7303/SYN69090156
(2025).“).
Data description
The final released dataset includes a total of 441 STL format files and corresponding NRRD format files along with 147 Case IDs. Specifically, this comprises 147 STL format files and corresponding NRRD format files of original defective mandibular models, 147 STL format files and corresponding NRRD format files of preprocessed mandibular models and 147 STL format files and corresponding NRRD format files of postprocessed mandibular models. Additionally, the dataset includes diagnostic results from oral and maxillofacial surgery experts regarding HCL classification of the defects, factors influencing reconstruction strategies, defect origins, and image quality assessments.
The STL format files and corresponding NRRD format files, as well as diagnostic results, serve as valuable resources for training deep learning frameworks to classify mandibular defect types and design reconstructions. The defect origin data may facilitate research into the risk factors and clinical challenges associated with reconstructing different types of mandibular defects. It should be noted that this dataset does not distinguish between training and testing subsets.
Data format
The directory structure and file types of the Mandibular Defect Dataset are shown in Table 3.
All 3D models are stored and named as “[HCL class]_[CaseID]_Original.[format]”, “[HCL class]_[CaseID]_Pre.[format]” or “[HCL class]_[CaseID]_Post.[format]”, while the STL files and corresponding NRRD files are respectively stored in directory “Cases_NRRD” and “Cases_stl”. The virtually reconstructed mandibular models, which serve as manual annotations for the defective models, were created by experienced oral and maxillofacial surgeons. “[CaseID]” is an integer representing the unique identifier for each case within the dataset. Additional detailed annotations and diagnostic classifications provide robust support for both clinical and research applications in mandibular defect reconstruction.
Technical Validation
The schematic diagram of the reconstruction design performed in ProPlan CMF 3.0 is shown in Fig. 6. A clinical CT axial slice from the case is presented, along with the original model reconstructed via thresholding, the preprocessed model after repositioning, and the final postprocessed model. To provide a clearer visualization, the reconstructed part is segmented using Boolean subtraction. Yellow represents the part of the maxilla, blue represents the part of the mandible, green represents the reconstructed part. For cases with complex defects, the annotated postprocessed models demonstrate good edge continuity and symmetry, with curvatures closely approximating those of a healthy mandible, thus meeting clinical requirements.
Fig. 6
Virtual reconstruction designs for different types of mandibular defects.
To further validate the annotated models, the data are imported into a commercial 3D printer (Objet260 Connex3, Stratasys Ltd., MN, USA) and printed using MED620 resin material (Stratasys Ltd., MN, USA), as shown in Fig. 7. The 3D printed models are used to visually demonstrate the manufacturability of the dataset, specifically watertightness, absence of self intersections, and absence of nonmanifold edges. The prints undergo independent review by two oral and maxillofacial surgeons, and the continuity of the reconstructed margins and the plausibility of the overall morphology are confirmed.
Fig. 7
The most representative 3D-printed resin models of L defect (a) and HCL defect (b). The original defect margins are marked with a black permanent marker on the postprocessed models.
Usage Notes
This dataset has a wide range of potential applications. Beyond its use for AI training as mentioned earlier, it can serve as a cohort study database for analyzing mandibular defects, such as exploring potential correlations between defect types, associated diseases, and the complexity of reconstruction. Additionally, the data can be utilized for anatomical analysis of deformed mandibles, whether using traditional image processing methods or neural networks.
While this dataset reflects the clinical distribution of mandibular defects, it has some limitations, primarily due to significant differences in the incidence rates of various subtypes. All defect types in the dataset represent acquired mandibular defects, and the dataset lacks sufficient data on rare subtypes, such as C and HH defects. As it is nearly impossible to collect pre-defect healthy mandibular data from patients with mandibular defects and associated diseases, the reconstruction process relies solely on manual annotations by clinical experts. This step is highly dependent on clinical experience and may not be applicable to all potential use cases.
Despite its limitations, this dataset holds significant value for training AI models in autonomous reconstruction planning for mandibular defects. One of the future objectives is to expand the dataset by incorporating additional data on rare mandibular defect types. Its utilization is encouraged, and proper attribution and citation are kindly requested when it is used.
Code availability
The version and parameter information of the relevant software and hardware used in the above workflow can be found in the Methods section.
References
Jewer, D. D. et al. Orofacial and mandibular reconstruction with the iliac crest free flap: a review of 60 cases and a new method of classification. Plast Reconstr Surg 84, 391–403; discussion 404-395 (1989). 1.
Chatterjee, D. et al. Reconstruction of complex oro-mandibular defects by four different modifications of free fibula osteomyocutaneous flap: A prudent alternative to multiple flaps. J Plast Reconstr Aesthet Surg 75, 3346–3355, https://doi.org/10.1016/j.bjps.2022.04.060 (2022).
Article PubMed Google Scholar 1.
Frensilli, J. A., Kornblut, A. D. & Tenen, C. Reconstruction of a mandible after shotgun trauma: report of case. J Am Dent Assoc 110, 49–51, https://doi.org/10.14219/jada.archive.1985.0287 (1985).
Article PubMed Google Scholar 1.
Xie, Q., Jiang, X. & Huang, X. Distraction osteogenesis application in bone defect caused by osteomyelitis following mandibular fracture surgery: a case report and literature review. BMC Musculoskelet Disord 25, 813, https://doi.org/10.1186/s12891-024-07922-z (2024).
Article PubMed PubMed Central Google Scholar 1.
Coletti, D. P., Ord, R. & Liu, X. Mandibular reconstruction and second generation locking reconstruction plates: outcome of 110 patients. Int J Oral Maxillofac Surg 38, 960–963, https://doi.org/10.1016/j.ijom.2009.03.721 (2009).
Article PubMed Google Scholar 1.
Goh, B. T., Lee, S., Tideman, H. & Stoelinga, P. J. Mandibular reconstruction in adults: a review. International journal of oral and maxillofacial surgery 37, 597–605 (2008).
Article PubMed Google Scholar 1.
Kumar, B. P., Venkatesh, V., Kumar, K. J., Yadav, B. Y. & Mohan, S. R. Mandibular reconstruction: overview. Journal of maxillofacial and oral surgery 15, 425–441 (2016).
Article PubMed Google Scholar 1.
Annino, D. J. Jr. et al. Accuracy and outcomes of virtual surgical planning and 3D-printed guides for osseous free flap reconstruction of mandibular osteoradionecrosis. Oral Oncol 135, 106239, https://doi.org/10.1016/j.oraloncology.2022.106239 (2022).
Article PubMed Google Scholar 1.
Ettinger, K. S. et al. Patient-specific Implants Improve Volumetric Surgical Accuracy Compared to Stock Reconstruction Plates in Modern Paradigm Virtual Surgical Planning of Fibular Free Flaps for Head and Neck Reconstruction. J Oral Maxillofac Surg 82, 1311–1328, https://doi.org/10.1016/j.joms.2024.06.166 (2024).
Article PubMed Google Scholar 1.
Hurley, C. M. et al. Current trends in craniofacial reconstruction. Surgeon 21, e118–e125, https://doi.org/10.1016/j.surge.2022.04.004 (2023).
Article PubMed Google Scholar 1.
Lo, L. J. & Lin, H. H. Applications of three-dimensional imaging techniques in craniomaxillofacial surgery: A literature review. Biomed J 46, 100615, https://doi.org/10.1016/j.bj.2023.100615 (2023).
Article PubMed PubMed Central Google Scholar 1.
Tatti, M. et al. Segmental Mandibulectomy and Mandibular Reconstruction with Fibula-Free Flap Using a 3D Template. J Pers Med 14 https://doi.org/10.3390/jpm14050512 (2024). 1.
Moiduddin, K., Anwar, S., Ahmed, N., Ashfaq, M. & Al-Ahmari, A. Computer assisted design and analysis of customized porous plate for mandibular reconstruction. Irbm 38, 78–89 (2017).
Moiduddin, K., Al-Ahmari, A., Nasr, E. S. A., Mian, S. H. & Al Kindi, M. A comparison study on the design of mirror and anatomy reconstruction technique in maxillofacial region. Technology and Health Care 24, 377–389 (2016).
Article PubMed Google Scholar 1.
Tantisatirapong, S. et al. The simplified tailor-made workflows for a 3D slicer-based craniofacial implant design. Sci Rep 13, 2850, https://doi.org/10.1038/s41598-023-30117-w (2023).
Article PubMed PubMed Central Google Scholar 1.
Heo, M.-S. et al. Artificial intelligence in oral and maxillofacial radiology: what is currently possible? Dentomaxillofacial Radiology 50, 20200375, https://doi.org/10.1259/dmfr.20200375 (2021).
Article PubMed Google Scholar 1.
Li, J. et al. Synthetic skull bone defects for automatic patient-specific craniofacial implant design. Sci Data 8, 36, https://doi.org/10.1038/s41597-021-00806-0 (2021).
Article PubMed PubMed Central Google Scholar 1.
Li, J. et al. Sparse convolutional neural network for high-resolution skull shape completion and shape super-resolution. Sci Rep 13, 20229, https://doi.org/10.1038/s41598-023-47437-6 (2023).
Article PubMed PubMed Central Google Scholar 1.
Mazzocchetti, S. et al. Neural shape completion for personalized Maxillofacial surgery. Sci Rep 14, 19810, https://doi.org/10.1038/s41598-024-68084-5 (2024).
Article PubMed PubMed Central Google Scholar 1.
Wodzinski, M. et al. Deep learning-based framework for automatic cranial defect reconstruction and implant modeling. Comput Methods Programs Biomed 226, 107173, https://doi.org/10.1016/j.cmpb.2022.107173 (2022).
Article PubMed Google Scholar 1.
Wu, C. T., Yang, Y. H. & Chang, Y. Z. Three-dimensional deep learning to automatically generate cranial implant geom