Embracing computer vision for diagnostic maxillofacial imaging — An artificial intelligence machine learning (AIML)pilot project

Abstract

Background: Artificial intelligence (AI) is rapidly transforming healthcare, particularly in diagnostic medical
imaging. For Nigerian Oral and Maxillofacial surgeons, embracing AI technologies is essential to improve
diagnostic accuracy and maintain global relevance. This study aimed to demonstrate the potential of machine
learning (ML) tools in enhancing diagnostic precision in maxillofacial radiology.
Methodology: A supervised learning model was developed using Google’s Teachable Machine, a no-code ML
platform based on computer vision. Radiological images of histologically confirmed lesions were retrieved. Two
projects were conducted: Project 1 trained the model to distinguish between malignant and benign bony jaw lesions
using 46 radiographs (panoramic and sectional CT images). Project 2 trained the model to differentiate between
craniofacial fibrous dysplasia and ossifying fibroma, using 40 radiographs. Each model was tested on five new
images. The output probabilities were analyzed, and standard performance metrics—accuracy, precision, recall
(sensitivity), and F1-score—were computed. Additionally, ROC-AUC (Receiver Operating Characteristic – Area
Under the Curve) curves were generated using Python code on Google Colaboratory IDE.
Results: In Project 1, the model yielded predictive probabilities ranging from 89% to 100% for distinguishing
malignant from benign lesions. In Project 2, it produced 71% to 100% probabilities for classifying fibrous
dysplasia versus ossifying fibroma. Applying a 70% probability threshold for positive prediction, both models
achieved perfect scores (1.0) across all performance metrics, including AUC = 1.00.
Conclusion: AI-driven computer vision models show strong potential for improving diagnostic workflows in
maxillofacial imaging. Their application can support more efficient clinical decision-making. However, the use of
small test samples may have resulted in overfitting. Future studies with larger datasets and increased AI literacy
among clinicians are essential for real-world implementation in resource-limited settings.