EasyCeph

Cephalometric analysis is the analysis of the dental and skeletal relationships of a human skull. Dentists and orthodontists use it as an aid for treatment planning. It involves manually tracing bone structures in an X-Ray and evaluating their relationships with each other.  The process

Project Title

EasyCeph

Project Area of Specialization

Artificial Intelligence

Project Summary

Cephalometric analysis is the analysis of the dental and skeletal relationships of a human skull. Dentists and orthodontists use it as an aid for treatment planning. It involves manually tracing bone structures in an X-Ray and evaluating their relationships with each other. 

The process can be divided into four steps:

• Manual tracing of the existing bone structure.
• Specific landmarks are identified on the diagram.
• Distances and angles between the landmarks are computed.
• An analytical approach is applied on the computed values to identify features of interest.

Dentists and orthodontists have reported that this process is time consuming and labor intensive. There is a high degree of human error, which can have a significant adverse impact on critical surgical decisions.

To resolve this problem, we will be creating a software that can automatically generate the cephalometric tracing of a given lateral cephalometric x-ray. We aim to build a program that would be able to detect landmarks with an accuracy like that of a well-trained medical professional. It is hoped that this program would significantly decrease the time required for the process. A single analysis could be completed in a few seconds which would be a significant improvement from the 10 minutes required for a manual analysis. We plan to study past research work and available methods/systems and their limitations. We will also investigate why the available methods/systems give unreliable results in some measurements. We aim to develop an improved method that will produce more accurate and more reliable results of landmark detection, especially in some specific landmarks and measurements where the existing systems produce unreliable results.

The program will utilize advanced computer vision techniques to accurately detect the landmarks. The model will be trained and tested on large sets of annotated x-rays. X-rays will be sourced from Pakistani medical institutions as well as online open-source datasets. Annotations for the x-rays will be provided by skilled medical professionals based in Pakistan. They will utilize commercial manual annotation software. To ensure the quality of the dataset, a peer review mechanism will be used to verify the annotations.

Project Objectives

To develop an AI-powered cephalometric analysis software that automatically generates the cephalometric tracing of a given lateral cephalometric x-ray. And to help practitioners complete single analysis in a few seconds, which would be a significant improvement from the 10 minutes required for a manual analysis.

Project Implementation Method

• The application’s frontend utilizes maps (dictionaries) with nested arrays to store landmark information.

• For storing datapoints, each datapoint is contained in a tuple where the first element in that tuple is an image in a NumPy array and the second element is a list of 19 tuples representing the 19 xy-coordinates' landmarks.
• For dataset representation, 3 objects have been prepared to point to train, validate, and test datasets.
• For image transformation, albumentations.ai library has been used. From this library, algorithms to apply resize, random flips, random 90-degree rotations, normalization, affine and elastic transformations have been used.
• For modeling, CNN with 4 layers was used, with the first 3 having convolutional, batch normalization, relu and max pool layers, and the last one being fully connected layer to produce an output of 19x2 dimension. And a pure ResNet50 model was implemented from scratch and applied to our problem.
• We trained ResNet50 on 20, 45, and 80 epochs. We saw reduction in the training and validation loss until we crossed 120 epochs where we saw no further significant improvement in the loss. For the 90 epochs case, train loss was: 6.357 while validation loss was 14.725.
• For optimization, Adam optimizer algorithm was used, and since it is an image regression problem, our loss criteria were Mean Squared Error loss.

Benefits of the Project

EasyCeph is not like any other AI tool you would see in the market; it aims in aiding orthodontists to provide hassle free landmark tracing. With EasyCeph, now doctors won’t have to manually trace the entire Ceph Xray to find landmarks. Our AI predictor predicts landmarks for you within seconds.

This process just takes three steps:

1. Adding a patient
2. Uploading their lateral cephalometric x-ray
3. Letting our model do the rest to get predictions

Now, with our fast AI landmark predictor, doctors can get Ceph Xray diagnosis more quickly and easily, freeing up their time to focus on more pressing areas of their practice. In addition to getting Ceph landmarks in seconds, we also allow doctors to adjust the predicted landmarks so that the final landmarks can be as accurate as possible by keeping doctors in the loop. With our tool, you'll get a more precise diagnosis thanks to a doctor's final check and the ability to adjust points.

Moreover, since our tool will be mostly used by medical professionals, we have ensured to keep the UI as simple and clean which have been praised by our beta testers doctors as well, thereby striking another important proposition of our product as it makes our AI tool user friendly, and doctors can easily adjust to this tool.

Technical Details of Final Deliverable

Tools:

• Scikit
• Google Colab for model development and data processing
• Flask/Django/Express.js for server-side development
• React.js/Flutter for web application
• TensorFlow/PyTorch for creating DL model.
• NumPy
• OpenCV
• Matplotlib

The application utilizes a micro-service architecture as it uses two different independent back-end servers. The Flask server is used for cephalometric predictions while the Google Firebase server is used for authentication and data storage.

The application’s front-end (client) is deployed on a remote server and is accessed from the end-user’s browser. There are two back-end servers. The Flask server is used for cephalometric predictions and the Google Firebase server is used for authentication and data storage.

The front-end application uploads cephalometric x-ray images to the Flask server which then returns the landmark predictions. The application then sends the predictions and annotated cephalometric x-rays to the Firebase server where they are stored on the Cloud Firestore database for future use.

The Cloud Firestore data storage service which comes under the umbrella of Google’s Firebase backend service is used for persistent data storage. It is a cloud hosted NoSQL database. User data including cephalometric x-rays, landmarks, account information etc. are stored in it.

The application utilizes the HTTPS protocol to transfer data between the client, model server and Firebase server. DNS (Domain Name System) protocol is used to match the URL of the deployed application with the appropriate IP address.

Final Deliverable of the Project

Software System

Core Industry

Medical

Other Industries

Core Technology

Artificial Intelligence(AI)

Other Technologies

Sustainable Development Goals

Good Health and Well-Being for People, Industry, Innovation and Infrastructure

Required Resources

If you need this project, please contact me on contact@adikhanofficial.com