Image Set with Labeled Segments
Console MQTT Messages
DermLens empowers patients with Psoriasis to better manage their condition by leveraging AWS DeepLens.
Psoriasis is a common chronic skin disease affecting as many as 7.5 million Americans and more than 125 million people worldwide. The disease can have an immense negative impact on people’s quality of life. Improving the management of the condition will have a big impact on individuals and society.
The most frequent symptoms of psoriasis are visual and can be captured by DeepLens. Psoriasis is characterized by patches of abnormal skin, typically red and scaly.
Using DeepLens, we can deliver a cost-effective, scalable solution to change the lives of millions by empowering patients with psoriasis to monitor and manage their condition. Running machine learning algorithms and inference locally in real time is a game changer, allowing classification and segmentation on the edge of the network at the location of the patient, rather than depend on high bandwidth connectivity to centralized hardware.
We used a dataset of 45 labeled images of skin with abnormal segments. Each image in the training set comes with a mask indicating the abnormal skin. We trained a model on DeepLens using the images in the training set.
Our cloud integration uses AWS. In addition to DeepLens, we used MQTT to enable our App to have a stream of information coming from the DeepLens device. Our lambda function is triggered to return a confidence estimate for the severity of the psoriasis, based on the segmentation and the percent of the image that is potentially seen as at risk. We used MXNet and Tensorflow for model training, along with hyperparameter refinement with Sagemaker. The model was evaluated by training both on-device (see the YouTube video) and by Sagemaker with our images in a S3 bucket.
In addition to AWS console and CLI resources, the project builds on node.js and Python Lambda functions, along with Python notebooks using Juypter and AWS resources for learning. We used a combination of two laptops, one desktop computer and AWS resources, along with DeepLens for the building and testing.
We created a companion mobile app for self-reporting of additional symptoms such as itching and fatigue, using ClojureScript and React. It is intended to be used in a Continuous Care scenario in which the reported data is available for the physician and care team. We envision that DeepLens ultimately can be embedded into a hardware product for psoriasis patients that also captures such additional symptoms.
Tom Woolf is a Professor of Physiology at John Hopkins University and the co-founder of DaiWare, a startup enabling patients to track and understand their health data.
Terje Norderhaug has a degree in Computer Science and is the co-founder of Predictably Well, a digital health startup empowering patients to better manage autoimmune conditions.