This project aim to detect 8 tooth and 4 gums diseases. It was solved using deeplearning with image classification and some diseases with segmentation. It was a project for Smiletronix a British startup that delivers AI-powered home dental diagnostics. I was the first AI engineer and build a dental dataset with the participation of dentists worldwide using a cloud labelling app. This dataset was double checked as we were looking for medical/dental associations approval.
The first prototypes were trained with tensorflow (object detection API) as we were exploring the possibilities of a future exportation to a tflite model for predictions on the edge. Models were learning acceptably and the main problem we found was on the dataset building as different doctors use to have different opinions on disease definition in the same image. Those are classical problems on medical dataset like imbalanced classes and large intraclasses correlation and big differences in the same clases (for example, different caries levels)
