Dental Open Bite Detection
I recently completed a deep learning project which involves detecting the presence of dental open bite using traditional convolutional neural network (CNN) and transfer learning so I thought I should write about it.
Overview
Dental open bite is a condition characterized by a lack of vertical overlap between the upper and lower teeth when the mouth is closed. Open bite can lead to various dental and facial issues, affecting aesthetics, speech and self-esteem — orthodontistwilmington.
The objective of this project is to develop a model that can accurately detect whether a dental image exhibits signs of open bite or not. By leveraging traditional convolutional neural networks (CNN) and transfer learning techniques, I aim to create a solution for automated open bite detection in dental images.
Dataset
Due to unavailability of open bite dental images, I manually gathered a dataset of 1000 dental images from various sources on the internet. The dataset consists of 350 images labeled as open bite and 650 images labeled as non-open bite. The images are in JPEG format and have varying dimensions.
Model
I built three different variants of CNN networks: a traditional CNN, an Inception v3 model, and a VGG16 model. The traditional CNN was constructed using multiple convolutional and pooling layers, followed by dense layers, and was compiled with a categorical cross-entropy loss function and Adam optimizer. Inception v3 and VGG16 models were utilized as well, leveraging their pre-trained architectures for improved performance on the new dataset. These models were chosen to explore different approaches and compare their effectiveness in the task
Result
The performance of the three CNN variants was evaluated using precision, recall, and F1 score metrics. The results are summarized in the table below:
Although these results indicate relatively modest performance across all models, which can be attributed to the limited size of the dataset and class imbalance, this project nonetheless provides a solid foundation for future research and development in automated open bite detection. Future improvements might focus on expanding the dataset to include a larger and more diverse set of dental images, which is expected to enhance the model performance metrics significantly.
