/chestxray.ai

Primary LanguageJavaScriptMIT LicenseMIT

ChestXRay AI

⚠️ ⚠️ DISCLAIMER: This is a project to demonstrate the usage of Javascript Tensorflow. Do NOT use this for REAL for Medical Diagnosis. ⚠️ ⚠️

This is a very simple page that lets you upload an X-Ray image of you Chest, and predicts the probability of having any of these 4 diseases:

  • Cardiomegaly
  • Mass
  • Pneumotorax
  • Edema

image

Privacy

X-Ray images are never uploaded to a server. Instead, all the calculations are done in the browser, meaning your X-Ray image never leaves your computer, and thus it is not shared with anyone else.

Credits

This project was initially developed at the 21st SUSE Hackweek project, and based on DeepLearning.AI AI for Medical Diagnosis course.

Thanks to SUSE for letting me hack on this project for a week. Thanks to DeepLearning.AI for such brilliant course, and the permission to use and publish the model for non-comercial purposes.

The CAM visualization was copied and modified from https://github.com/tensorflow/tfjs-examples/tree/master/visualize-convnet.

Dataset

The dataset used is the result of this research paper: Xiaosong Wang, Yifan Peng, Le Lu, Zhiyong Lu, MohammadhadiBagheri, Ronald M. Summers.ChestX-ray8: Hospital-scale Chest X-ray Database and Benchmarks on Weakly-Supervised Classification and Localization of Common Thorax Diseases, IEEE CVPR, pp. 3462-3471,2017@InProceedings.

The dataset is available at the NIH Clinical Center, America's Research Hospital

This dataset has been annotated by consensus among four different radiologists for 5 of our 14 pathologies:

  • Consolidation
  • Edema
  • Effusion
  • Cardiomegaly
  • Atelectasis

Model

This is using the saved model that was trained as an exercise in the first week of AI for Medical Diagnosis. Then, it performs the same transformations that were done during training to an X-Ray Image: scaling and normalization, and computes a prediction. All this is done with javascript, meaning that we are using the hardware (CPU/GPU) from the user.

To save the model I had to add this to the coursera assignment: model.save("models/nih/saved_model.h5")

To transform the model from keras to tensorflowjs:

docker run -it --rm -v `pwd`:/python evenchange4/docker-tfjs-converter tensorflowjs_converter --input_format=keras /python/save_model.h5 /python/saved_model.tfjs

This model was implemented and trained by DeepLearning.AI.

Model Metrics

See https://github.com/jordimassaguerpla/model.chestxray.ai/blob/main/metrics.md for the metrics.

Deploying

Since we are using a saved model with tensorflowjs, we do not need a "big server", but just the web server of your choice (i.e. apache2) to serve the static files. This could also easily be distributed with a Content Delivery Network (CDN).

Just copy the model from https://github.com/jordimassaguerpla/model.chestxray.ai/tree/main/saved_model.tfjs into the webserver, and then copy the files from this project.

Legal

Except for the model, you can mostly do what you want (MIT license). For the model, see https://github.com/jordimassaguerpla/model.chestxray.ai/blob/main/README.md#legal