Nowhere is as much attention paid to the ability of AI to change our lives as in healthcare. In the Dell EMC AI Innovation Lab, we have developed AI models to classify 14 different thoracic pathologies using a dataset of 120,000 frontal chest x-ray images released by the National Institute of Health. We have created the necessary code and artifacts to both perform training and use inferencing workloads on Apache Spark, traditional CPU clusters, and GPU accelerated servers.

The following examples use the NIH dataset and chest x-ray model:

• Training and inferencing on an Apache Spark cluster
• Training on NVIDIA V100 GPU, inference with NVIDIA T4 GPU
• Training and inferencing on CPU
• Flask Web App demo for use case

The following figure shows the AI-assisted radiology application that we published through Domino Data Science Platform. The application detects diseases in a selected chest x-ray and also provides probabilities of the occurrence of 14 diseases. Domino Data Science Platform provides usage metrics, access control, and other capabilities for managing the application.

Figure 9.          AI-assisted radiology as a published application in Domino Data Science Platform

We found Domino Data Science Platform well suited for this use case. The ability to use all three types of compute—CPU, GPU, and distributed CPU—in a single platform allowed us to run the workload on a different platform when all the resources in one of our compute pools was in use.

For examples of our code for this use case, see BigDL-ImageProcession-Examples on GitHub. We have also published a whitepaper in collaboration with Intel about the chest x-ray use case with the Analytics Zoo library to enable training and inferencing on Spark.