Abstract
Background: Odontogenic cysts are common jaw lesions that can significantly affect oral and overall health if not diagnosed and managed promptly. Among them, Odontogenic Keratocysts and Dentigerous Cysts are particularly noteworthy due to their distinct biological behaviour and treatment approaches. Odontogenic Keratocysts are known for their aggressive nature and higher recurrence rates, making accurate differentiation from Dentigerous Cysts critical for appropriate clinical management. Histopathological examination remains the gold standard for diagnosis; however, it is time-consuming and subject to interobserver variability. The integration of deep learning with digital pathology presents a promising solution, offering enhanced diagnostic accuracy, consistency, and efficiency in identifying and classifying these lesions. Objective: To develop and evaluate a deep learning-based method for differentiating Odontogenic Keratocysts and Dentigerous Cysts using whole-slide histopathology images. Methods: Formalin-fixed, paraffin-embedded, Haematoxylin and Eosin-stained histopathology slides were digitised at 20× magnification using a Morphle whole-slide scanner. A total of 110 whole-slide images (65 Odontogenic Keratocysts and 45 Dentigerous Cysts) were tiled, resulting in 684 annotated image tiles (345 Odontogenic Keratocysts, 339 Dentigerous Cysts). Various deep learning models, including pre-trained and custom CNN architectures, were trained and evaluated. A novel CNN model, OdonGAP, incorporating four convolutional layers, three dense layers, and a Global Average Pooling (GAP) layer instead of a Flatten layer, was developed to improve feature extraction and classification. Results: The OdonGAP model achieved a testing accuracy of 73.72 %, with optimal parameters including 64 filters, kernel size 3, learning rate 0.001, 20 epochs, and batch size 16. The use of the GAP layer enhanced feature representation by effectively reducing spatial dimensions. A user interface, DeepClassify, was also developed to support automated classification in clinical settings. Conclusion: This study demonstrates the potential of deep learning and digital pathology in differentiating odontogenic cysts, offering a step toward improved diagnostic accuracy and clinical integration. Future work will focus on expanding the dataset and evaluating additional models to further enhance diagnostic performance.