The management of an impacted mandibular third molar (M3M) is a commonly encountered yet intricate challenge in oral and maxillofacial surgery. Population studies across different countries suggest that most young adults must decide whether to have surgery to remove third molars or retain and monitor until problems arise¹. Decisions regarding surgical management depend on a careful assessment of anatomical structures, such as the mandibular canal (MC), adjacent molars, and alveolar bone, as well as their spatial interrelationships^1,2. Accurate evaluation is critical: misjudging the degree of root proximity to the inferior alveolar nerve (IAN) or overlooking subtle anatomical variations may increase the risk of nerve injury and postoperative complications^3,4,5. In this context, reliable and explainable diagnostic tools can help dentists and specialists weigh multiple factors and tailor their approach to each patient’s unique anatomy, ultimately improving treatment outcomes and patient satisfaction.

Artificial intelligence (AI), particularly deep learning (DL), has shown substantial promise in medical imaging tasks, including the segmentation of complex structures in cone-beam computed tomography (CBCT) scans^6,7. In oral and maxillofacial imaging, DL-based methods have demonstrated remarkable accuracy in delineating teeth and the MC^8,9, potentially aiding surgical planning. However, these purely data-driven models often function as “black boxes,” offering limited transparency regarding the reasoning behind their outputs when clinical decisions are needed in high-risk situations^10,11. This lack of interpretability is detrimental to clinical adoption, as surgeons must understand the rationale behind a given recommendation to confidently integrate AI insights into their decision-making processes.

While existing explainable AI (XAI) techniques — such as saliency maps, feature attribution, and post-hoc interpretability frameworks — have improved transparency to some extent^12,13, they may not fully reflect the logical, stepwise processes clinicians use to evaluate surgical risks and plan interventions. Notably, oral and maxillofacial surgeons often employ sequential reasoning, first interpreting anatomical features, then assessing spatial relationships, and finally applying established clinical criteria to inform treatment decisions^1,2,5. Recognising this, recent advances in AI propose reasoning modelling strategies, in which the decision-making pathway is segmented into modular steps, each informed by the preceding output^14,15. Such approaches parallel the cognitive processes of experienced clinicians and may better bridge the gap between model predictions and surgical reasoning.

Within this paradigm, hybrid models — DL-driven image analysis integrated with rule-based decision criteria in a logic layer — offer a particularly appealing solution¹⁶. By encoding clinically validated guidelines and risk assessment protocols into the logic layers of an analytic pipeline, these systems can provide explanations that align with the factors surgeons consider. For instance, DL models can detect and segment relevant anatomy, but they often fall short of translating these findings into direct, actionable guidance for clinical decisions. A logic layer, encoded with explicit rules, can quantify the minimum allowable distance between the MC and the tooth root for safe extraction or identify specific angular relationships associated with increased procedural complexity. This can directly benefit resident and early-career surgeons who incur higher rates of postoperative complications and nerve injuries in M3M surgery¹⁷. This structured incorporation of domain knowledge can foster value to AI-generated detections, ensuring that the machine-generated outputs not only improve diagnostic accuracy but also translate into auditable guidance on safer decisions potentially leading to improved patient outcomes. Indeed, emerging evidence in other areas of oral and maxillofacial practice suggests that blending machine intelligence with human-like reasoning can reduce iatrogenic complications and streamline clinical workflow quality^18,19.

Therefore, this study aims to investigate if a fully explainable automated system can be developed that simulates the decision-making process for the surgical management of M3Ms, as observed on CBCTs. While our study utilises labels for the automatic segmentation of key anatomical structures, we aim to generate clinical decisions that do not rely on expert-labelled data. We hypothesise that a hybrid diagnostic pipeline, which couples a robust DL-based segmentation with an innovatively designed 3D spatial analytics module, namely logic layer, can generate reliable high-risk decision recommendations. This module leverages advanced 3D data processing techniques, combined with a robust suite of spatial distance-based rules rigorously aligned with established surgical criteria, to deliver enhanced diagnostic precision and clinical applicability. By channelling the predictive power of AI through a decision sequence, the system generates interpretable recommendations for managing M3Ms. We validate its performance against expert assessments of the same CBCT scans, thereby evaluating its diagnostic reliability and the clarity and relevance of its reasoning.

The study was approved by the Ethical Committee of the Commonwealth Scientific and Industrial Research Organisation (CSIRO) under the approval number 2022_040_LR. It was performed in accordance with the Declaration of Helsinki. Informed consent was obtained from all participants.

2) Per-tooth structure segmentation (crown and root) using a customised model from the nnU-Net framework²¹, a U-Net-based implementation tailored for medical imaging with automatic architecture optimisation²².

A total of 84 scans were utilised for training and validation of Mandible and MC segmentations, and 8 scans were used for testing. An optimised nnU-Net architecture was configured to target specific functional groups relative to the Mandibular Bone and MC. Specialised preprocessing, including image standardisation and rescaling, was applied to the input image, followed by a 3D U-Net to detect the trabecular bone as a reference landmark. Based on the position of this reference landmark within the input image, a distinct crop was generated for each landmark group and subsequently segmented with a dedicated 3D U-Net. The purpose of this approach is to reduce the input size to the 3D U-Net for enhanced efficiency. Following the segmentation, the results underwent post-processing, including artifact removal and smoothing. Figure 1 illustrates the workflow of the CBCT segmentation process.

Data augmentation strategies were employed during training to expand the dataset and enhance its diversity, thereby improving the model’s generalizability and robustness.

The output of the system includes voxel masks representing the segmentation results, providing a visual representation of the identified structures within the CBCT images. Furthermore, the system generates STL mesh files, offering a three-dimensional representation of the segmented dental anatomy. Figure 2 illustrates the segmentation output from a representative CBCT scan.

1b. Consider coronectomy due to the complex relationship between the M3M and the IAN. Horizontal impaction increases surgical complexity and the risk of nerve injury during crown sectioning. Sectioning the coronal tissue in multiple planes is advised²⁴.

The final output of the decision was calculated by a conditional function following the decision criteria, illustrated by Fig. 7 below.

Our results demonstrate that the association of DL segmentations with rule-based spatial analysis can generate reliable recommendations, offering a robust and interpretable alternative for clinical decision-making in M3M management. Although expert-labelled data were used for the development of segmentation models, to the best of our knowledge, this is the first study to generate AI-based simulations of clinical decisions in M3M surgical assessment without using expert annotations.

Recent literature has shown several applications of hybrid rule-based DL tools in healthcare¹⁶. Specifically, spatial analysis coupled with DL segmentations can increase interpretability without sacrificing accuracy^28,29. MacCormick et al.²⁸ produced a 15° cup-to-disc rim profile for glaucoma, achieving high discrimination using a dataset 100 times smaller than typically required for DL methods. In urology, a 3D ultrasound system automatically segmented anterior urethral strictures; its measurements of stricture length and fibrosis matched those from intraoperative and manual assessments, and the images were readily interpretable to surgeons²⁸. As demonstrated by the results of our study, these examples illustrate how thoughtfully designed AI can augment clinical insight by presenting outputs in human-interpretable formats. Such technology enables decision support that clinicians can validate and act upon with confidence.

In the context of M3M surgery, our spatial analysis strategy aims to leverage robust and promptly assessable DL-based segmentation models of maxillofacial structures^6,30, while addressing interpretability limitations noted in prior studies. Our segmentation results have met, or surpassed performance of multiple other studies as demonstrated by recent literature reviews and experimental designs analyzing the same structures^6,8,9. Regarding treatment decisions, recent studies using DL models have achieved impressive accuracy in tasks such as automatically classifying impaction type and assessing extraction difficulty^31,32. For example, Balel et al. integrated the classic Pell & Gregory, Winter, and Pederson schemes into a single YOLOv11-based model, enabling automated classification of impaction position and difficulty with over 95% accuracy, comparable to expert judgment³¹. Similarly, CNN approaches have been used to predict M3M extraction difficulty scores (e.g., Pederson index) from panoramic radiographs, showing that AI can mirror expert assessments to a high degree³². However, these end-to-end models function primarily as black boxes – they output a classification or score without explaining the anatomic basis for the prediction. Consequently, even when models achieve impressive performance metrics, such as 95% accuracy, clinicians cannot verify how the AI arrived at a conclusion, whether it is correct or not. Such limitation undermines trust and hinders clinical adoption. In contrast, our strategy focuses on using DL in a time-consuming and lower-complexity task, namely, the segmentation of key anatomical structures, while delegating the higher-complexity task of clinical decision-making to spatial analysis. In line with our methodology, Carvalho et al. used a pipeline of DL models to detect the MC and classify the degree of third-molar root overlap in panoramic imaging¹⁸. Their proposed system achieved high accuracy and produced a diagnostic table to recommend when CBCT should be obtained, based on combinations of nerve overlap and root development status. We expand on this idea by exploring findings in cases where prescription of CBCT was necessary, specifically, after initial screening with conventional radiographic methods indicated a complex and specific clinical question²³. By leveraging spatial analysis, our approach addresses three critical limitations of studies in the field: it offers interpretable outputs in textual and visual forms; it minimizes the requirement for extensive datasets comprising high-quality expert-annotated data by encoding rule-based logic; and it addresses the unavoidable class imbalance caused by limited samples in specific clinical conditions (such as the prescription of CBCT in M3M assessment). Our design strategy not only benefits M3M assessment but also has the potential to support decision-making in other oral surgery applications that heavily depend on 3D and 2D imaging, such as orthognathic surgery¹⁹, as well as areas of clinical dentistry that also depend on diagnoses and treatment planning based on dimensional relationships of anatomical structures of known morphological variability. These include applications in orthodontics, prosthodontics, and periodontics, which could use similar spatial analysis rules organised in different configurations.

Previous studies have highlighted imaging features, such as narrowing of the canal lumen and the canal positioned in a bending or groove within the root complex²⁵, as significant aspects of M3M management. Such features were observed in a few cases from our clinical evaluation sample; however, this has not influenced treatment planning decisions in our study, as all such cases presented in conjunction with verification of direct contact between the M3M and the MC. While we understand that surgical management of M3Ms needs to take into consideration analysis of components beyond simple detection of diagnostic imaging features, such as patient age, complex medical history, and patient preference²⁶, our study has focused on the key aspects driving clinical decision, as supported by recent literature in the field of M3M management based on CBCT imaging ^20,23. Nonetheless, the development of a comprehensive AI-based system for M3M management, taking into account medical history and patients' preferences, seems achievable as of the time of manuscript submission. For that, further research may explore the potential of an HDLRB tool integrated with agentic hierarchical systems^33,34, as well as perspective studies targeted at clinical outcomes, particularly in cohorts of junior clinicians or early career specialists who can potentially translate the assistance provided by comprehensive AI-based systems into safer clinical practices.

There are several limitations to our study. Cases with M3Ms presenting with caries lesions within dentine, evidence of immature root development, or evidence of apical inflammatory lesions had to be excluded from our validation sample. Future systems may integrate AI systems that provide reliable detection in such cases⁷, as these features determine different surgical treatment planning strategies. Our clinical validation sample size is relatively small, particularly in cases where M3Ms were fully impacted (n = 4). Obtaining larger sample sizes is challenging, as the indication for CBCT should be reserved only for cases where initial screening with conventional radiographic methods indicated a complex and specific clinical question²³, which has been demonstrated in a study where surgical treatment plans only changed by CBCT in 12% of the time²⁵. Larger studies are needed to verify the reliability of the system's outputs. Nevertheless, the crucial element of our design is its ability to enable clinicians to critically analyse the output decision in an easily interpretable format, thereby offering reassurance when the system's recommendations differ from those of dentists and specialists.

Further investigation into the ethical ramifications of AI tools, such as the HDLRB tool, is warranted. Although a recent checklist for AI in dentistry addresses principles such as transparency, diversity, privacy, equity, solidarity, and governance, additional research is necessary concerning the pillars pertinent to clinical applications³⁵. These encompass wellness, respect for autonomous decision-making, accountability, responsibility, prudence, and sustainable development^35,36. As suggested by Balel et al., regionally limited datasets may affect the performance of AI models in assessing third molars in underrepresented populations³¹. Although all subjects investigated in our study come from a single country, the ability of our system to maintain robustness and generalizability depends solely on the segmentation model's capacity to identify anatomy, a prospectively favourable scenario. Therefore, by providing clinicians easily explained auditable outputs, this study offers an alternative to increase trustworthiness, usability, and transparency in AI applications in healthcare, as also highlighted by relevant working groups³⁷.

In conclusion, our study demonstrates that an HDLRB tool can effectively simulate the clinical decision-making process for the surgical management of M3Ms as observed on CBCTs without the need for expert-labelled data. This alternative offers a fully interpretable automated solution that addresses clinicians' concerns with ‘black box’ predictions in high-risk situations, therefore facilitating clinical adoption.

E.L.D., X.L., K.L., and S.L. conceptualised the study. E.L.D., M.H., X.L., J.D.O., Sh.L., S.K. and C.S. designed the study. E.L.D, M.H. and K.R. performed the clinical evaluation. E.L.D. performed the statistical analysis. E.L.D., X.L., and J.D.O. analysed the data. E.L.D. and X.L. participated in writing the original draft of the manuscript. All authors revised and approved the final version of the manuscript to be published. D.W. and H.S. provided resources and supervised the study.