Page 90 - AIH-1-3

Artificial Intelligence in Health Dental cavity prediction with computer vision

Applying these augmentation techniques expanded
the dataset to a total of 2383 images, thereby facilitating a
richer and more comprehensive training process.

3.1.3. Manual annotation process
Following the augmentation phase, we proceeded with the
manual annotation of dental cavities within the augmented
images. For this purpose, we employed the Roboflow
annotation tool, which facilitated meticulous annotation
of dental cavities (Figure 2). The chosen method for object
detection involved bounding box annotation, represented Figure 2. Annotation using the Roboflow annotation tool
by a rectangular box icon within the annotation tool. In the
annotation process, annotators utilized crosshairs to determine A
the starting point for drawing bounding boxes around dental Utilized 218 smartphone
cavities. Each bounding box served as an annotation for the images of dental cavities. Optimized parameters
presence and location of a dental cavity within the image. of the YOLOv5 model
Furthermore, the Class Selector within the tool allowed Expanded the dataset As an additional measure
of severity, cavity masks
annotators to assign the appropriate label to each annotated to 2,383 images by image from SAM were evaluated.
augmentation.
bounding box, signifying the presence of a dental cavity. This the area of each bounding box
Based on the applied model,
manual annotation process was performed for approximately Trained the YOLOv5 model was determined, to estimate
400 images, ensuring the availability of accurately labeled data on the training set. the severity of the cavity.
for the subsequent training of the object detection model B
focused on dental cavity identification. Take several photos of teeth. YOLOv5 model estimates:
Focus on teeth that may
have cavities. • Number of cavities
3.2. Dental cavity localization using YOLOv5 • Location of cavities
• Severity of cavities
The schematic of the research project is shown in Figure 3A. Upload images to
Once the predictive model is developed, its application website or app. Information is kept in
is straightforward. Implementation can be developed the app or is sent to a
dental professional
for a smartphone app, where images would be taken by YOLOv5 model evaluates for treatment.
the images.
patients or dental assistants, without the need for a dental
professional. Figure 3B describes how the prescriptive Figure 3. (A) The schematic of the research project. Training set will
model would be used. be trained and optimized by YOLOv5 Model. SAM (Segment Anything
Model) will be evaluated. (B) The use of the prescriptive model. The
3.2.1. Introduction to YOLO number of cavities, locations and severity of them will be estimated.
Abbreviation: SAM: Segment anything model.
In our pursuit of precise dental cavity localization with
the augmented dataset, we harnessed the power of the a systematic process aimed at enabling it to accurately
YOLO object detection framework. YOLO represents predict the presence and location of dental cavities within
a groundbreaking approach to object detection, our annotated images. Our training dataset, enriched by
characterized by its remarkable speed and accuracy. Unlike augmentation techniques and manual annotations, was
traditional object detection models, YOLO processes used to train the YOLOv5 model. This training process
images in a single pass, making it especially efficient for involved iteratively fine-tuning the model’s parameters and
real-time applications. The YOLO algorithm divides optimizing its ability to recognize dental cavities in varying
an image into a grid and predicts bounding boxes and image contexts. The YOLOv5 model’s training process was
associated class probabilities for each grid cell. This unique rigorous, ensuring a high degree of accuracy and robustness
methodology enables YOLO to excel in scenarios where in detecting dental cavities within the images. The successful
objects of interest may vary in size and scale, making it training of this model constituted a crucial milestone in our
precisely suited for our task of dental cavity localization. endeavor to automate dental cavity localization, facilitating
more precise and efficient dental health assessments.
3.2.2. Training YOLOv5 model
The YOLOv5 model, an evolution of the YOLO 3.2.3. Bounding box prediction
architecture, served as the cornerstone of our dental cavity Following the successful training of the YOLOv5 model,
localization efforts. Training the YOLOv5 model involved we transitioned to the crucial phase of bounding box

Volume 1 Issue 3 (2024) 84 doi: 10.36922/aih.3184