HeadHunter

Image classifier's confusion matrix
Heatmap generated by model
Uncertainty classification provided by model
Receiver operating characteristic curve (area = accuracy)
Project Poster

Inspiration

Modern day processes for detecting neurodegenerative diseases require expensive procedures, such as doctors appointments, CT scans (Computed Tomography Scan), MRI scans (Magnetic Resonance Imaging scans), and psychiatric meetings. Although these methods are effective at finding neurodegenerative diseases, they are very costly and are not viable solutions for people of low equity. These methods also do not extend to people in 3rd world countries, where these procedures may not even exist. As a result, hundreds of thousands of people in rural areas lack proper diagnosis, leading to almost one fourth of children under the age of eight with neurodegenerative diseases such as autism to go undetected and untreated. Not only this, therapists hold subjective views based on experience, race, gender etc., which can interfere with the autism diagnosis. As a result, we aim to optimize autism detection models in order to increase accuracy and decrease loss and false negative predictions, as well as implement explainable autism detection through heat maps and attention mechanisms. We used ensemble learning and uncertainty quantification to combine and optimize multiple models in order to increase the accuracy of our model. We also used a novel method of improving saliency maps called attention feature emphasis in order to explain where it detects abnormalities, which current ML models do not do. These heatmaps will allow parents and therapists to gain an objective view on points of interest for autism detection as well as point out possible things that they may have missed during diagnosis. After testing, we were able to increase accuracy by over 20%, decrease loss by over 30%, decrease false negatives by more than 30%, and minimize false negatives to 3%. We also were able to eliminate noise and focus on specific features that heavily influenced the model’s decision. In conclusion, due to the increasing number of undetected autism, we were able to create and optimize current autism models as well as implement high precision explainability. This system can easily be implemented into a web application and can be used around the world.

What it does

This project will provide patients with a more cost effective solution to accurately finding and diagnosing neurodegenerative diseases without the need for expensive scans or therapy sessions. This project allows people with low income to still be able to effectively and objectively detect neurodegenerative diseases in loved ones. It will also highlight the reasons that it came to the conclusion of the neurodegenerative disease being present or not.

How we built it

This project uses several numpy, pandas, tensorflow, and sklearn libraries to build the model

Challenges we ran into

Implementing GRADCAM was very difficult and caused us many issues. Reading images and formatting them was also very difficult due to the variation in images from our dataset.

Accomplishments that we're proud of

We achieved a very high accuracy for our image classifier.

What we learned

We learned how to implement explainability into machine learning models.

What's next for HeadHunter

We want to provide our models with a much larger dataset to prepare the model for extremes in the data. We also want to perform cross validation on the entire dataset rather than a small set so that we can better prevent overfitting. Lastly, we want to fine tune some of the layers of the IMAGENET base we used to improve the feature extraction.