crypto-doctor

Logo (intermittent)
Data Flow

CrypDoc adresses:

Lack of privacy for important medical data.
High friction in crowdsourcing with doctors.
Lack of incentive to share and contribute to research.

README: https://github.com/diffusioncon/Team-32/blob/master/README.md

Inspiration

When you had a MRT scan the scans are analyzed by a professional doctor to predict certain diseases, this can take a huge amount of time, thus waiting for the patient, as well as uncertainty due to only one oracle.

With our solution the MRT scan gets encrypted locally and send to a client which, first, runs an encrypted machine learning model on it to make a initial guess (predocted diagnoses) plus its confidence.

To ensure the machines prediction a bounty is associated to the task of prediction and spread to doctors around the world. The bounty is based on the confidence of the model, the higher the confidence the smaller the bounty.

The doctors can accept the challenge and try to solve the task. If a sufficient amount of experts agree on a result, without knowing the others decision, the diagnoses will send back to patient. In parallel, the experts get their bounty plus a raise of reputation. With a higher reputation you are more likely to get higher bounty-tasks.

Meanwhile the pure meta data of the date and the final diagnoses will be added to a ocean protocol data set, which can be bought be researchers and insurances to adjust their approach of business.

What it does

PreDoc is a platform for the remote diagnosis of medical conditions by a decentralized network of doctors.

How we built it

We trained multiple neural networks using facebooks pytorch and its corresponding encryption library to run encrypted models on encrypted data. For training data we used the Kaggle Brain MRI Images for Brain Tumor Detection and trained an addapted LeNet model for fast inference time, additionally we trained a Resnet18 for higher accuracy.

A simple mock up to demonstrate a possible user interface, here: https://balsamiq.cloud/s3vqw2q/pm6kxa1/r2278

Challenges we ran into, Accomplishments that we're proud of, What we learned

utilizing cutting edge Multiparty Computation (MPC) for non-trivial image classification
devising a locally running architecture that combines normal web technologies with pytorch distributed communication for MPC

What's next for crypto-doctor

payment system for patients with Coinbase Merchant, IOTA or similar
consensus among multiple doctor decisions with Enigma
training on encrypted image data: active learning with consensus doctors as oracles