Taking advantage of work distribution upon doctors. In a capital city, the capacity of doctors can quickly be overloaded but on the country side at the same time it might me the other way around. We implemented a solution to decentralize work and share tasks cleverly while being secure.
In addition, we create an incentive for additional diagnoses and the exchange of results of expertise and for research
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.
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 challenge and trained an addapted LeNet model for fast inference time, additionally we trained a Resnet18 for higher accuracy.
You can train your own model using:
python enc_train.py \
--source-dir-train ../brain-mri-train \
--source-dir-eval ../brain-mri-val \
--training-run-out models/resnet18.pth \
--backbone resnet18 \
--max-epochs 5if you add the argument --validate-encrypted True it will validate the trained network with the plain model and
the encrypted model with encrypted data
Once the seed model exists, you can use it to make predictions on encrypted images and test an interactive loop between patients, model and doctors.
Starting crypto-doctor:
python predictor
This will launch two services, one for patients and one for doctors. The program prints out information from both services prefixed by their process rank.
In the default configuration, you can interact as follows.
As PATIENT, post your image file to: http://localhost:8080/image e.g.:
curl -X POST -F "file=@test/Y22.jpg" localhost:8080/image
At no point in time, the automatic prediction has access to the original image.
As DOCTOR, make your final decision at: http://localhost:8081/decision
In order to share the stream of final decisions, forward stdout to a seashell.
python -u predictor | nc seashells.io 1337
The resulting .csv is live at the seashells url printed at the beginning. We
published an example stream to Ocean commons.
A simple mock up to demonstrate a possible user interface, here
- 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
- a great deal of theoretical insights
- a great deal of practical insights (thanks for support and implementation from FAIR/CrypTen people)
- payment system for patients with
Coinbase Merchant,IOTAor similar - training on encrypted image data (tracking issue)
- consensus among multiple doctor decisions with
Enigma
- love 💚
- caffein ☕
- python 🐍
- pytorch 🔥
- crypten 🔒
- ocean protocol, squid-py 🦑
- fetch.ai/Enigma for mapping out next steps