/CML-Classifying-Melanoma

Primary LanguagePython

Deep Learning In Medicine

Classifying Melanoma on Cloudera Machine Learning

Summary

In this demonstration we will build a clissifier for melanoma, using a VGG16 Convolutional Network, and transfer learning.

Excercises

  1. Take open source images of skin lesions, and use those to build a classifier to detect malignant skin lesions
  2. Evaluate the performance of the model using TensorBoard, and matplotlib in CML
  3. Deploy the model onto a mobile device for use in clinical settings
  4. Use the mobile app to determine if a patient needs critical attention from a physician.

In the demonstration we use a model deployed on a mobile device. I.e. our inference happens on the edge, using a MobileNet model. The more likely choice for this use case would be to perform point inference by calling a more accurate (and compute intensive) model hosted centrally, or to perform classification in batch using a more accurate (and compute intensive) model.

1.

3.

4.

Talk Tracks (Preliminary):

Deck:

Demo Setup

The setup takes 5 minutes

1.

In CML Go to Projects, and create a New Project

2.

Name the Project "Melanoma Classification", and in the initial setup use git repo: https://github.com/hortonworks-sk/CML-Classifying-Melanoma.git, and hit the create button

3.

Launch a Python 3 workbench session

4.

Navigate to the load-libraries.sh script, and run the script. This will load the libraries needed for the demo.

5.

Navigate to the start_tensorboard.py script, and run this.

6.

Check that the Tensorboard link is displaying in CML and that tensorboard is running, by clicking the tensorboard link

7.

Click on the tensorboard tabs for Graph and Histograms, to check that these are displaying correctly.

8.

Navigate to experiments and click run experiment

9.

Run experiments for the _Inception3.py , and ** _VGG16.py, scripts. Use the python 3 kernel. No need to supply arguments for these.

10.

When these runs have completed, you should see the experiments listed as successful in the experiments view (as in the screenshot below)

Use Case & Industry Applicability

  • Use Case: Diagnosing Melanoma

  • Broader Healthcare Applicability:

    • Disease diagnosis using medical images
      • radiology (arteriography, mammography, radiomics)
      • dermatology
      • oncology

  • Broader Industry applicability

    • Biotech
    • Pharma