Inspiration
Oral diseases are known as the “neglected epidemic” because it is a widespread problem that has garnered little attention over the past decade. Many oral diseases progress in stages and can be prevented and reversed with early detection and simple treatments.
What it does
AlphaDentist is a preventative care tool directed towards elderly patients. Our easy-to-use online platform reliably detects gingivitis (initial stage of gum disease). Such a triaging tool would be especially useful in nursing homes and in correctional facilities, where structural injustices allow institutions to keep patient maintenance costs low by ignoring patients’ oral health. In correctional facilities, it is not uncommon for there to be 1,000+ inmates per dentist, and oral health is especially poor given that dental floss is prohibited and considered a security threat. In nursing homes, dental neglect is common as some residents may have difficulty voicing their concerns and because high costs of dental care would come out of nursing home's margins. In short, our screening tool aims to address dental neglect.
How we built it
AlphaDentist contains 3 sections.
Dataset Due to the privacy issues, majority of the healthcare dataset are not opened to public. We managed to generate our own dataset, with each of the data coming from reliable references. Gingivitis has diverse stages, from an initial stage where it's impossible to discern with human eyes, to an acute stage right before advancing to periodontitis. Later stages are easily distinguishable due to explicit symptoms such as bleeding or swelling of the gingiva, while earlier stages are not. Thus, we made sure to collect the images of gingivitis and healthy gum that are hardly distinguishable.
Each of the raw 'Gingivitis' dataset and 'Normal' dataset consists of 21 images. To increase the number of the data and to enhance the robustness and generalization of the machine learning model in terms of low-quality photos, various data augmentation methods were applied. Affine transformation, gaussian blur, gray scale, and three different color-jittering methods were used. Final datasets consisted of 147 images each, with 70% of the data used for training, and rest for testing.
Deep learning model Convolutional neural networks (CNN) have led to remarkable breakthroughs in image classification. Among them, residual network (ResNet) designed by Kaiming He outperforms most of the co-existing CNNs. We used ResNet34, which consists of 34 layers. Best model trained for 100 epochs shows 97.508% of accuracy on the test dataset. Grad-CAM, a stream of Explainable Artificial Intelligence (XAI), which visualizes the local region of the image where the model recognizes as the feature for the classification, proves that our model genuinely learn the different features between healthy gingiva and gingiva with gingivitis.
Website The website is the key to connect the client with our product. We used IceWrech's Flask-React-Boiler-plate as a reference to connect the React frontend website with the Flask backend server. The website includes a minimal form that allows users to upload their oral image, and our deep learning model will return their percent likelihood of possessing gingivitis within seconds.
Challenges we ran into
Connecting the React frontend with Flask server In order to connect the machine learning output with the website, several challenges were encountered. This include changing the whole directory to fit the webpack format and using Flask as the connection between the two different platforms. As this was our first time connecting both platforms together, we spent a lot of time debugging and researching this hurdle.
Scarcity of dataset There were not as many images of gingivitis and healthy gum available in the internet as we would like. Initially, we implemented an automatic image crawler to download the large number of images, but most of the collected data were inadequate. Thus, we manually examined each image thoroughly by searching for image on reliable websites and identifying if it was properly labelled, among other criteria.
Non-uniformity of dataset Unlike public datasets designed for deep learning tasks, such as CIFAR10 or Caltech-UCSD Birds, images pulled from internet varied in size. In order to train these images with neural network, they must have equal sizes. For overall consistency, we added padding pixels to mitigate the information loss occurred by cropping and cropped to (224*224) sized images.
Did the model truly learn the knowledge Compared to other machine learning tasks, our problem suffers from lack of data. This indicates that even though the accuracy for the test dataset is high or training and testing loss are gradually decreasing, the model may not truly understand the underlying principles of the classification. Thus, it is extremely important to manually analyze if the model has learn the proper knowledge. We tried number of different methods, such as adding Squeeze and Excitation Network (SENet) module or adjusting normalization function for better generalization, and finally determined experimentally through GradCAM that it is functioning well.
Accomplishments that we're proud of
Based out of the United States, Hong Kong, and South Korea, the three of us are proud to be an international team with diverse backgrounds and skill sets: biomedical engineering, front-end development, and machine learning development. This being the first healthcare hackathon for all three of us, we were proud to be able to apply our skill sets to a healthcare problem that has largely been neglected.
First approach to design ML model for gingivitis classification As far as we know, there is no previous work on gingivitis classification via machine learning. Despite of limitations on dataset building as well as time for implementing advanced deep learning model, our model performs 97.508% on the test dataset, and 90% on the unseen validation dataset. If we can get more reliable data and design the model based on dental knowledge on the symptoms of gingivitis by collaborating with dentists, we can outperform our current version of the model.
What we learned
Given the time zone difference and the fact that we each tackled a different component of the project, we relied heavily on effective communication. We learned so much from each other in such a short period of time and also received invaluable feedback from various mentors. This allowed us to examine our product not only from a technological and medical perspective, but also from a business and investor perspective.
What's next for AlphaDentist
Our next steps would be to partner with a team of experienced dentists to bring our product to reality. We would like to continue to develop our gingivitis classification model and add a few additional features to our website. We would love to see AlphaDentist grow into a start-up company and make a mark in the teledentistry industry. We stand by our personal mission of using technology to make healthcare more accessible to all.
Log in or sign up for Devpost to join the conversation.