Automatic Detection of Dentigerous Cysts on Panoramic Radiographs: A Deep Learning Study

Introduction

• Background: Odontogenic cysts and benign odontogenic tumors are usually painless and asymptomatic unless they grow large enough to cause significant issues. These lesions can often be identified through routine radiographic examinations.
• Problem: Accurate diagnosis of these lesions requires radiographic interpretation training and experience.
• Solution: Convolutional Neural Networks (CNNs) are increasingly used in medical imaging to assist in diagnosis. This study aims to develop a model for detecting dentigerous cysts on orthopantomographs (OPGs) to introduce dentistry students to AI applications.

Materials and Methods

• Data Collection: Two 5th-year dentistry students identified 36 OPGs with histopathologically confirmed dentigerous cysts.
• Image Processing: Images were resized to 1024x514 pixels, and augmented with vertical and horizontal flips for training-validation.
• Model Training: A U-Net CNN model was trained with 200 epochs using PyTorch, with a dataset split into 112 training images and 16 validation images.
• Evaluation: The model's performance was tested with new OPGs and evaluated for precision, sensitivity, and F1 score.

Results

• Performance Metrics: The model achieved a precision of 0.5, a sensitivity of 1, and an F1 score of 0.67.
• Example Detections: Figures show successful detection of dentigerous cysts attached to mandibular third molars.

Discussion

• Challenges and Limitations: The study's small sample size and the exclusion of cases without histopathological confirmation limited the model's sensitivity and accuracy. Public datasets are needed for broader testing.
• Comparison to Other Studies: Similar studies have achieved varying levels of success in detecting odontogenic lesions, with some reaching high sensitivity and specificity.

Conclusion

• Findings: The CNN model demonstrated potential for detecting dentigerous cysts even with a small dataset.
• Future Directions: Larger datasets and improved methods could enhance model accuracy, providing valuable diagnostic support for new dentists.

Acknowledgements and Contributions

• Contributions: E.O. and I.T. detected and segmented the cystic cavities, supervised by G.U. I.S.B. and O.C. processed the images and created the model.
• Conflict of Interest: None declared.