top of page

Artificial Intelligence in Dentistry

With new technologies and artificial intelligence many industries are improving their digital workflows. Today, the world of dentistry is clearly among those industries. Day after day, practices are changing through the use of new digital and artificial intelligence tools. The dentist improves his treatment techniques by making them more precise; the patient journey is even more transparent, secure and efficient.


In this article, we present a project aiming to develop a computer aided diagnostic tool for dental imaging, using the latest advances in Deep Learning applied to Computer Vision.


The company AIDEN - specialized in remote dental consultation assisted by AI - collaborated with MCOVISION in order to get support and expertise on :

  • collection and annotation of dental radiographs

  • implementation of Deep Learning algorithms applied to detection and segmentation of several anatomical and pathological structures

  • the deployment of APIs in production for use of the algorithms in real conditions


Artificial Intelligence and Computer Vision


Artificial Intelligence, or AI, is a field of research mimicking the behaviour of the human brain, to accomplish a certain number of tasks. When data are images or videos, we talk about Computer Vision.


Computer Vision is therefore a subfield of AI that allows computers to extract, analyze, and interpret the information contained in an image or video. Convolutional neural networks, or CNNs, used in Computer Vision (e.g. [1]) have made it possible to achieve never before seen levels of performance.


Medical imaging: a challenge for AI


Medical imaging is one of the most complex fields of imaging, due to its own constraints:

  • the acquisition and annotation of medical images can be complex, variable (noise, machine bias, ...), requiring a lot of expertise

  • medical data is very sensitive and therefore must be anonymized and secured

  • structures to be detected or segmented in medical images are sometimes rare and not very observable in practice, as is the case for certain pathologies; the data are therefore sometimes few in number, which handicaps the neural networks during learning; we talk about imbalance data

To overcome those challenges, having experience in image processing, and more specifically in medical image processing, is essential. Our consultants cultivate this very specific expertise, and use their experience to accelerate the AI health projects of our customers.


Computer-aided diagnostic system for dental radiography


Aided diagnostic is a task performed by many systems using Computer Vision. In dental radiography, this is an essential task that every dentist is required to perform frequently. Even for the best practitioner, the daily workload can sometimes lead to misdiagnosis. Practices and levels of expertise can also vary from one clinic to another, making it difficult to coordinate radiographic analysis and interpretation between different practitioners.


Thus, in order to improve the diagnosis in dental radiography, we have developed an artificial intelligence capable of describing a radiography, that is to say of automatically detecting and locating the anatomical and pathological structures inside the image.


The implementation of such a computer-aided diagnostic tool took place in three stages:

  • the annotation of radiography images, i.e. manual contouring of areas of the image to be detected or segmented; this first step is essential to allow artificial intelligence to learn correct and precise information

  • training an object detection model to locate teeth, several dental components and pathological structures on dental radiographs

  • the training of a semantic segmentation model aimed at delimiting contours precisely on radiographs of several anatomical and pathological structures

Finally, the combination of predictions from detection and segmentation models will allow the tool to offer a first diagnosis to the dentist.


Note that before communication of the AI resulting predictions to the patient, the results must be validated and if necessary corrected by the dentist.




Supervised Deep-Learning


Computer-aided diagnostic that we have developed is based on training of Deep-Learning models. These algorithms, based on neural networks, learn to predict from an annotated image dataset whose ground truth is known. Thus, in order to be accurate algorithms, an annotated image dataset of several thousand images is required.


From a mathematical point of view, training a neural network consists in back-propagation within the network of the prediction error in order to improve the prediction. For each a new image or batch of images fed into the network, the prediction is compared to the ground truth annotation. If it differs too much from the ground truth, the network adjusts its weights to better perform the next iteration. In other words, it is by comparing the prediction with the ground truth that the network learns. We talk about supervised learning. To have a robust model, the network must train on a large number of rigorously annotated images. Once trained, the model can then predict on unannotated data, and be deployed in real conditions. Thus, the quality of the data is essential and greatly influences the performance of the algorithms.


Once trained, the model is able to predict on new images never before seen by AI in a few seconds. We talk about inference. Pathologies and anatomical structures can be detected and will be used to propose an adequate treatment plan for the patient.


AI and the role of doctors


It is important to emphasize that industrialization of new AI algorithms in no way ignores the practitioner. On the contrary, it serves to assist him in his decision-making. The final diagnosis is up to the dentist, but he will have been able to benefit from a “first reading” beforehand, which will allow him in the long term to save time and mitigate his considerable workload. AI is actually the best ally of doctors, who will produce thanks to it a more precise diagnosis, in a reduced time.



References


[1] Ronneberger, Olaf and Fischer, Philipp and Brox, Thomas. U-Net: Convolutional Networks for Biomedical Image Segmentation. 2015




25 views0 comments

Recent Posts

See All
bottom of page