Odontogenic tumours (OT) are heterogeneous lesions derived from the tooth-forming apparatus (1–3). Although these tumours are relatively rare, comprising of less than 5 % ofall oral and maxillofacial biopsy specimens (4–7); they encompass a diverse spectrum of pathologies, ranging from hamartomatous to locally invasive and neoplastic lesions (1–3).
The World Health Organization (WHO) published the classification of (OT) in 1971 (8), with multiple revisions till the most recent 5th edition in 2022. In 1992 calcifying odontogenic cyst was introduced as an OT (9) and in 2005 odontogenic keratocyst (OKC) was reclassified as keratocystic odontogenic tumour (KCOT) (10), due to its locally aggressive nature and links to the PTCH1 gene and SHH signalling pathway (11, 12). However, they were subsequently reclassified as cysts in the 4th edition in 2017 (13) and in 2022, citing a lack of compelling evidence for neoplastic behaviour ((11, 12).
Ameloblastoma and odontogenic keratocyst (OKC) are common types of odontogenic lesions and particularly noteworthy due to their potential for locally infiltrative behaviour and the risk of recurrence (14, 15). The recommended treatment for conventional Amelobalstoma is resection of the jaw with a margin (16). While such treatments often necessitate intricate reconstructive surgeries for aesthetic and functional restoration, early and precise diagnosis can allow for more conservative treatments (17). Approaches such as enucleation and peripheral ostectomy become viable, which is crucial for paediatric patients to minimize disruptions to craniofacial development and emerging dentition (18, 19). With OKC on the other hand more conservative enucleation is a common treatment with adjunctive measures like Carnoy’s solution, 5FU, or cryotherapy employed to reduce the chances of recurrence (20–22).
A practitioner's provisional diagnosis and impression of a lesion are pivotal, setting the course for the patient's initial management, be it monitoring, biopsy, surgery, or referral to a surgical specialist. But for this methodology to be effective, the practitioner's initial clinical assessment must be accurate to minimise the possibility of over or undertreatment. Diagnostic delays can escalate disease progression, necessitate more intrusive surgeries, and dampen treatment outcomes. Thus, gauging the precision of clinical diagnoses against the definitive histopathologic diagnosis—the gold standard—is crucial. The clinicopathological concordance (CPC) quantifies this alignment, measuring the concordance between a clinician's preliminary diagnosis and a pathologist's conclusive pathological diagnosis of the biopsied specimen (23).
General dental practitioners (GDPs) often serve as the primary point of contact for many patients, equipped with training to inspect, diagnose, and handle a diverse range of oral ailments. In contrast, specialists undergo additional postgraduate training in their specific domains. When it comes to the CPC rate the literature presents varied conclusions. While some studies suggest specialists exhibit higher CPC rates, others contend that the outcomes between GDPs and specialists are comparable (23, 24). Reported CPC rates among specialists typically range from 50–70% (23, 25, 26).
Even with the result of the initial biopsy, securing an accurate provisional diagnosis is still paramount, as there is potential margin of error and hence discrepancies between incisional biopsy findings and the comprehensive specimen (27). For instance, Chen et al. highlighted a discrepancy rate of 11.1% between preliminary and definitive histopathological assessments. In such instances, clinicians might opt for an additional biopsy or seek a second opinion from another pathologist (28).
However, humans are prone to diagnostic errors stemming from human cognitive biases and heuristics can be minimised with the assistance of AI tools. Mental shortcuts, such as confirmation bias and the availability heuristic, often lead clinicians to make judgments based on initial impressions or memorable experiences(29–31). AI, with its data-driven approach, can provide a more objective analysis, free from personal biases. By offering evidence-based suggestions, AI can challenge and expand a clinician's differential diagnosis, countering biases like anchoring and overconfidence. While human expertise remains paramount, integrating AI can serve as an additional check, enhancing the accuracy and consistency of clinical diagnoses.
Artificial intelligence (AI) is machine-based simulation of human intelligence which has started to integrate into daily lives and into healthcare (32). AI has advanced clinical applications by enhancing patient results, making processes more efficient, and cutting costs. In the clinical realm, AI has made remarkable strides in tasks such as analysing data for image segmentation, aiding in clinical decisions like predicting disease outbreaks, and executing intricate procedures including surgeries and rehabilitation(32). This underscores its transformative potential for healthcare. In the field of oral and maxillofacial surgery, convolutional neural networks have demonstrated improved efficacy in identifying and categorising maxillofacial fractures, TMJ disorders, classification, and detection of malignant disorders (32–39).
The use of AI in the diagnosis of cysts and tumours of the jaw is an emerging field with promising potential. AI algorithms can potentially assist healthcare professionals in analysing medical images, identifying patterns, and providing diagnostic support. Previous studies have shown promising results in respect to identifying and classifying lesions in plain radiographs and cone-beam computed tomography (CBCT) scans (38–41). By leveraging machine learning techniques, notably deep convolutional neural network (CNN), AI systems can learn from large datasets of annotated images to recognise specific radiographic features associated with different types of cysts and neoplasms (38–45). Further to this by combining radiological and demographics and clinical data, AI models can generate differential diagnoses to provide clinicians with additional decision support(38–41) and for the purpose of surgical planning ((46–48).
Much of the research based on AI and CNN has been restricted to solely image-based methods, curtailing the effective relay of information, and not fully harnessing the potential of AI in clinical setting whilst Natural language processing (NLP) using large language model (LLM) excel in correlating textual and visual data, aiding in the interpretation of radiographs.
NLP is a form of AI that plays a key role in clinical decision support (CDS) systems (49). Generative Pretrained Transformer (GPT) model is one of such that is an open source available to the public. Such AI system were utilised as triaging systems and virtual clinics during COVID-19 pandemic (50, 51). Furthermore, NLP systems that extract disease symptoms from clinical texts have been used in identifying cofounding characteristics of patients in medical records and predicting disease outcomes (52, 53).
AI as CDS has been proposed to increase efficiency and accuracy of clinical diagnosis and therefore safety for the clinicians (49). Diagnostic accuracy has been tested in internal medicine based on common chief complaints where CDS have shown high rates of accuracy for achieving diagnosis (49). Currently there is a lack of study examining the ability for CHAT-GPT to assist clinician with diagnosis of jaw lesion and in particular OT and OKC.
The University of California Los Angeles developed the Oral Radiographic Differential Diagnosis (ORAD) system in 1990s. ORAD employs the Bayesian approach and Bayesian belief networks (BBN), tools that have seen applications in both medical and dental fields to enhance diagnostic accuracy (45, 54–57). The essence of this approach is that it can estimate the probability of a disease from a particular set of observed variables. This estimation is possible when the relative prevalence of each disease is known, along with the likelihood of the occurrence of the findings. A Bayesian Belief Network (BBN) showcases the numerical relationships between various nodes. Importantly, it avoids cyclical logical connections and clearly defines the directional relationships between these nodes. Utilizsing a directed acyclic graph, the BBN delineates the causality flow among the nodes. Furthermore, it quantifies the strength of connections between variables. An essential feature of a BBN is its capability to autonomously update probability estimates as new data is introduced.
The study aimed to investigate diagnosis range, relative frequencies, and clinical presentations of odontogenic cysts and tumors in a New Zealand population over a 15-year period. The study also explored CHAT-GPT and ORAD as a possible tool to assist with diagnosis relating to odontogenic jaw pathology. In this regard, the study sought to compare the CPC achieved by CHAT-GPT, ORAD, and by clinicians.