Deeplasia Service

Deeplasia is a prior-free deep learning approach to asses bone age in children and adolescents. This repository contains a RESTfull service with a simple API to process X-ray images and predict bone age in months.

Please refer for more information:

• http://www.deeplasia.de/
• https://github.com/aimi-bonn/Deeplasia

How to Use

In order to run this application, you must provide the deep learning models. Please contact use to get them.

Use the environment variable DEEPLASIA_THREADS to limit the number of threads used by PyTorch (defaults to 4 threads).

Run in Conda Environment

Requirements:

• Conda must be installed
• Deep learning models are located in the directory ./models

Run the following CLI commands and navigate to http://localhost:5000/.

conda create -n flask_ba python=3.9
conda activate flask_ba
pip install -r requirements.txt
python flask run 

Run with Docker

Requirements:

• Docker must be installed
• Deep learning models are not included in the image and must be mounted on container start

You can use our pre built Docker image to run the application:

docker run -p 8080:8080 -v ./models:/app/models ghcr.io/crescnet/deeplasia-service

Or you can build the image yourself (clone this repository first):

docker build -t deeplasia-service .
docker run -p 8080:8080 -v ./models:/app/models deeplasia-service

Navigate to http://localhost:8080/

Limiting CPU usage

To limit the CPU usage of the docker container, add the following flags to the docker run cmd:

--cpus=<number_of_cpus>

Note, that this should match the number of threads specified with environment variable DEEPLASIA_THREADS.

e.g.:

docker run -p 8080:8080 --cpus=2 -e "DEEPLASIA_THREADS=2" -v ./models:/app/models ghcr.io/crescnet/deeplasia-service

API

Please refer to deeplasia-api.yml for an OpenAPI specification of the API.

Request

In python the request can be conducted as follows:

import requests

url = "http://localhost:8080/predict"

test_img = "/path/to/xray.png"
files = { "file": open(test_img, "rb") }

data = {
    "sex": "female",  # specify the sex as "female" or "male"
    "use_mask": True,  # default is true
    "mask_crop":  1.15,  # default is 1.15
    "use_invChecker": True, # default is true
}

resp = requests.post(url, files=files, json=data)
resp.json()

Gives something like:

{
    "bone_age": 164.9562530517578,
    "used_sex": "female"
}

Usage of Masks

Skipping the masking by the predicted mask is meant to be a debugging feature, if the results with the mask are not convincing (e.g. predicting 127 months as age), one could re-conduct bone age prediction without the mask and see if makes a difference. We might think about storing the masks as a visual control as well as logging features in general in the future.

License

The code in this repository and the image deeplasia-service are licensed under CC BY-NC-SA 4.0 DEED.