@spu314

Welcome to my second devlog for SPARC!

–Finished Phase—————–
I just finished training the AI model to accurately identify points along the spine of a human on an image of their back. Originally, I planned to use the Scoliosis.v5i.yolov8 dataset, which I obtained from Roboflow; however, I had to switch to a much larger and more effective dataset that had ~400 images for training.
In total, I trained the larger dataset a total of 6 times on my CPU, with varying amounts of epochs each time.
Overall, this training phase required a lot of experimentation with the number of epochs (times the model goes through the entire dataset), as I was looking for high accuracy, but wanted to avoid overfitting. In particular, I didn’t want the AI model to train on the data too many times as it would eventually start memorizing the dataset instead of actually learning to identify the points.
In addition to training the model, I also tested it with a total of 30 images after each training session. During these testing phases, I discovered that the model works the best on images with sufficient lighting and relatively large image size.
Although this phase was sometimes frustrating as I had to train the model over and over again based on the results I obtained from testing it on images, it was extremely fufilling as I was able to achieve an Pose mAP50 of 0.84, which is better than I had expected after having so many failed attempts!

–Next Phase———————
Next, I will create a website that will allow users to test my model with images of their own, and potentially, their camera.
To add on to the website’s features, I will create a Math formula to classify the results of the image as normal, mild scoliosis, moderate scoliosis, or severe scoliosis.

–Image Description–––––––
The attached image shows one of my best results of the model on an image I found off the internet.

Hi! My name is Sophia, and this is my first devlog for SPARC!

––My overall plan for this project–––––––––––
My plan for this project is to train an AI model to be able to predict whether or not a person has scoliosis, and the severity of the scoliosis based on a picture or video of their back.
After training the model, I would like to create a website that people can visit to measure the severity of their scoliosis with simply their webcam.
In the future, I would like to track the measurements in a log and create a graph that will show the changes in the scoliosis measurements.

—Current Phase+Progress———————————–
Right now, I am working on training the model, testing it, and analyzing its accuracy. In particular, I want the model to find the coordinates of 10 points along the spine that I can use to calculate the deviation and the severity of the scoliosis.
I first learned about how to use YOLOv11n, a pretrained model, from https://docs.ultralytics.com/tasks/pose#train.Afterwards, I downloaded a dataset that the AI can train on from Roboflow.
Then, I set up and activated a virtual environment to prevent dependency conflicts, and installed ultralytics to use YOLO and PyTorch.
After coding up a python program, I was able to train the model for 20 epochs (each one complete pass through the dataset) on a very small dataset of ~50 training images. When I tested the model on an image, its predictions were inaccurate and had low confidence levels. It even detected the person’s back multiple times and made multiple predictions…
I realized that this dataset was too small, so I switched to a larger dataset of ~400 images, and am continuing to train my model until it is able to accurately detect the points along the spine.