Artificial Intelligence in dental health care – a case study

Dentistry is one of the oldest branches of medicine, its roots long predating agriculture. The earliest evidence of dental treatment comes from the upper paleolithic period. Like all medicine, dental healthcare today looks nothing like it did during the stone age. Machine learning, among our era’s most promising contributions, is a major factor in the difference. Dental health is among the most important areas of modern healthcare. Oral diseases are the most common non-communicable diseases in the world. According to WHO data, an estimated half of the global population struggle with dental problems like tooth decay. The most common diseases are tooth decay and periodontal disease. Tooth and gum treatments happen also to be costly, accounting for 5% of total health expenditures and 20% of out-of-pocket expenditures.

How do AI and healthcare come together?

According to Accenture, automated image diagnosis could be worth up to a $3 billion chunk of the healthcare market. Considering the state-of-the-art image recognition solutions currently available, the most obvious applications include MRIs and other scan interpretations. [image-comparator left=”/wp-content/uploads/2019/11/4small.jpg” right=”/wp-content/uploads/2019/11/4smallpredsbig.jpg” classes=”hover”][/image-comparator] Move your mouse cursor over the image to see the difference. According to estimates from IBM, up to 90% of all medical data is reported in image form. Thus, improving diagnostics and treatment with image recognition tools is the top-of-mind application. In healthcare, however, the road from idea to implementation can be far more tortuous than in other industries.

Challenges for AI in healthcare

Supporting modern healthcare is imperative to the quality of life humans can have. At the same time, the potential for machine learning technology to create breakthroughs that will benefit humanity is enormous. This makes it all the more important that the multiple challenges we face be overcome.

• The challenging process of obtaining data – because medical data is sensitive, it is not easy to create a database that contains sufficient entries that are also anonymous.
• The need for strict human supervision of artificial intelligence. There is no place for mistakes in healthcare – it is all a matter of human life and health. Also, there is no legal framework governing decisions made by AI and applied in treatment, so human supervision is key.

Keep smiling

Our client is a leading producer of dental equipment, delivering a full range of instruments – from single-use tubes and appliances to head and tooth x-ray devices. The x-rays their machines deliver are crucial in diagnosing multiple oral afflictions, including tooth decay, periodontal disease and multiple others. Yet the raw images x-rays present, enhanced only with slight processing and basic clarifications, is often all that the physician gets. Unlike with MRIs and the process of interpreting them, x-ray machines themselves don’t support the diagnostic process. [image-comparator left=”/wp-content/uploads/2019/11/1small.jpg” right=”/wp-content/uploads/2019/11/1smallpredsbig.jpg” classes=”hover”][/image-comparator] Move your mouse cursor over the image to see the difference. Our client was looking for a way to support dentists in their work by preprocessing x-ray images and providing more information about the potential diseases they turned up. The company also needed to automate the dull and repetitive work of spotting and writing up information in medical documentation. That information mainly concerned particular teeth depicted in the images their machines produced.

The teething problem

The initial goal was to properly mark each tooth in an image. This would prove more difficult than you might imagine, especially given the number of teeth that can be missing. In fact, according to the World Dental Federation, up to 30% of the population, aged 65-74 have no natural teeth left. This was reflected in the data the client provided, which we used to train our model. An average healthy adult human has 32 teeth, but there were an average of 26 teeth depicted in the pictures we dealt with and included in our model. The model was responsible for spotting the right positions of particular teeth, including telling which ones were missing.

Additional challenges included:

• Localizing implants, restorations and other things used in medical treatment
• Uncovering signs of the most common oral diseases such as cavities and periodontitis

This task was further complicated by the fact that cavities are so indistinct in x-ray images that an untrained eye has little to no chance of finding them (in contrast to distinguishing between Aliens and Predators). And if a random person struggles to see something, an algorithm will, too. From a machine learning point of view, the project was an object detection problem, similar to others we have handled during projects for Nielsen and the National Oceanic and Atmospheric Administration. We were given around 1000 anonymized, high-resolution x-ray pictures along with bounding boxes around teeth and other findings created by professional dentists and radiologists. [image-comparator left=”/wp-content/uploads/2019/11/3small.jpg” right=”/wp-content/uploads/2019/11/3smallpredsbig.jpg” classes=”hover”][/image-comparator] Move your mouse cursor over the image to see the difference. We used PyTorch’s MaskRCNN neural network model to simultaneously localize teeth and other findings and determine their labels. The model was supported by our custom-made algorithm, which checked the correctness of the order of tooth labels and, more importantly, canceled some predictions. A prediction might need to be ruled out if, for example, a neural network were to spot 17 teeth (where only 16 are possible in a healthy human jaw). In any case, this bonus step was only needed for a few images with multiple chained crowns inserted to replace a row of missing teeth. For the vast majority of images, the neural network predicted the tooth labels perfectly. We also implemented a set of heuristic postprocessing procedures addressing human knowledge not available to a model based on images only. Examples of human knowledge included:

• You can’t have a cavity outside a natural tooth: on your collarbone, on an implant replacing a missing tooth, on a crown, etc.
• You can have either a crown or a dental seal on your tooth.
• Even if you have a double root canal treatment in a premolar or molar tooth, the model should highlight it with a single frame in order not to confuse the dentist.

The accuracy

Our model found it easy to spot and identify particular teeth adhering to the World Dental Federation Notation, even for ‘unusual’ patients, such as those with only a few teeth left. It localized and labeled teeth, making only a handful of mistakes while processing hundreds of images. Also, most object categories, including implants, dental work and prosthetics, were detected properly, with the mean average precision more than 0.9 on a 0-1 scale. Finally, the model’s hyper-sensitivity ensures that the early signs of decay or other oral diseases are spotted. [image-comparator left=”/wp-content/uploads/2019/11/2small.jpg” right=”/wp-content/uploads/2019/11/2smallpredsbig.jpg” classes=”hover”][/image-comparator] Move your mouse cursor over the image to see the difference. When dealing with cavities, providing the user with more false-positives rather than overlooking the signs of a developing illness was considered a more feasible strategy. Images have to be reviewed by a dentist anyway, so delivering more warnings, even with a higher rate of false-positives, can deliver a great dose of support for dental professionals.

Summary

Using deep learning-based AI systems in healthcare is no longer confined to the realm of science fiction. In this case, a mere x-ray has armed dentists with images enhanced with ready-to-go information on the placement and state of every tooth in a patient’s mouth, as well as a cheat-sheet on the early signs of disease that could have been overlooked. This project is a perfect example of how such systems can reduce the dull, repetitive and mistake-prone work usually done by professionals, who could be putting their expert knowledge to better use.