Skip to content

thomas-pipithkul/CS410-VideaHealth

 
 

Folders and files

NameName
Last commit message
Last commit date

Latest commit

 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 

Repository files navigation

VideaHealth logo

VideaHealth is a Boston-based startup that brings AI to the field of dentistry by automating the process of dental x-ray analysis. Today, denticians have to spend countless hours analyzing these images and they still miss up to 50% of pathologies in dental x-rays. VideaHealth is helping dentists identify diseases and communicate treatment recommendations to their patients.

How to use:

  1. Install the dependencies
    pip install -r requirements.txt

  2. Run the main file
    python main.py

  3. To see the visualizer, add the command line argument -V or --visualize
    (Press 0 on your keyboard to cycle through the images)
    python main.py -V

Project Description

Group members: Tony Chau, Jeffrey Garcia, Cameron Hayes, Ke Liu, Paul Maynard, Danley Nemorin, Suncharn Pipithkul, and Jacob Janak.

We worked on VideaHealth's neural network as part of our CS 410 Intro to Software Engineering class in the Spring of 2021. The existing neural network they created was already outputting lots of predictions. Our job was to refine these predictions into a single result. This is how our project fits into VideaHealth's system:

Work flow

This is what the raw data that we were getting from the neural network looks like on a dental x-ray:

Raw data

In the above image, you can see that there are far too many boxes for the number of teeth. The nueral network is outputting every possible box that it can think of. Although this is not shown in the image, every box has a predicted tooth label and a probability score. We used these boxes and their respective labels and scores to trim the output down to one box per tooth. This is what the results look like after our postprocessing algorithm is run:

Our outputted results

You can see that each tooth now only has one box. It also has a label and a confidence score. We used a non-maximum suppression algorithm to correctly output the boxes. Non-maximum suppression is a common technique to filter the predictions of object detectors. We also wrote a seperate algorithm to assign the correct numbers to each tooth. We had to work around edge cases for people with missing teeth or other irregularities. This chart shows how teeth should be numbered:

Tooth numbering chart

Our results were good, achieving both a precision and recall of about 0.93. Precision shows how trustworthy each of our outputted boxes are and recall shows how many boxes are missing in our output. We were able to calculate these numbers by comparing our results to the ground truth data, which was determined by an actual dentist. The fact that our results were so close to the dentist's results is promising.

NOTE: We are not allowed to share all the x-ray data that we used because it is confidential medical information. We are only able to share a small sample from that data set.

About

No description, website, or topics provided.

Resources

License

Stars

Watchers

Forks

Releases

No releases published

Packages

No packages published

Languages

  • Python 100.0%