Page 92 - AIH-1-3

Artificial Intelligence in Health Dental cavity prediction with computer vision

The data in this study are based on image augmentation
imitating multiple possible alternatives for each
original image. While we believe that this is an accurate
representation of future images that will be available,
it would be more effective to have actual images from
multiple angles for teeth and cavities.
More broadly, as AI is applied to more diverse Figure 6. Source image and segmented image
opportunities in modeling and medical diagnostics, several
issues may emerge. These relate both to the development
of new models and the use of automated diagnostics by
individuals and medical professionals. On the modeling
side, one can foresee, in the not-too-distant future, the
possibility of automated modeling being employed,
evaluating a broad set of models on a given dataset. Without
supervision and effective parameter tuning, these methods
could lead to overfitting or the use of inappropriate models.
Similarly, the data used for these automated studies could
be suspect. Using available images that are not evaluated
by people could be unreliable. Imagine a situation where
an autonomous model is developed by images created by
AI, for example.
The implications of AI on medical practice should also Figure 7. Mask image of the cavity
be considered. Applications like the one proposed here
provide effective but limited self-diagnosing opportunities The SAM is also highly efficient, making it suitable for real-
to individuals, especially in areas with limited access to time applications. It can generate a segmentation mask for
health. However, it is likely that some people who could any prompt in real time after precomputing the image
receive effective diagnoses from medical professionals embedding.
would also use these tools. Given the noticeable rate of For instance, given a bounding box around a dental
false negative results, these individuals may not receive cavity in an image, the SAM can be used to segment only
the necessary treatment. There are also implications for that area using the SamPredictor class. The SamPredictor
medical professionals. As medical professionals become class takes a bounding box as input and outputs a
more reliant on technology, there is a risk of decreased segmentation mask for the cavity that is enclosed by the
expertise in the profession. This decline in expertise may bounding box (Figure 6).
arise from becoming overly reliant on forecasting tools or
from outsourcing diagnostics to low-cost services that rely 4.3. Dental cavity masks from SAM
on technology.
Image masks are binary images that represent the
4.2. Image segmentation using the segment foreground pixels of an object. These masks were used to
anything model (SAM) analyze the shape of the dental cavities inside the bounding
The SAM is a promptable segmentation system capable of box. Using the SAM, a segmentation mask for the cavities
zero-shot generalization to unfamiliar objects and images was generated. The segmentation mask was then cropped
without the need for additional training. This capability to the bounding box and analyzed to measure the desired
allows SAM to segment objects into new images; it has properties of the cavities. For example, the area of the
never seen before simply by providing it with a prompt cavities can be measured by counting the number of white
such as a text description, a bounding box, or a few clicks pixels in the cropped segmentation mask.
on the image. The white pixels in the binary-masked image show the
SAM is trained on a massive dataset of over 1 billion exact shape of the cavity (Figure 7).
segmentation masks, making it the largest segmentation 5. Conclusion
dataset to date. This extensive training allows SAM to
learn a wide range of object appearances and relationships, In this study, we embarked on a journey to revolutionize
enabling it to generalize to new images with high accuracy. dental cavity analysis, resulting in a holistic framework that

Volume 1 Issue 3 (2024) 86 doi: 10.36922/aih.3184