AI Radiology — Chest X-ray Screening System for Medical Organizations

With new advancements in AI, X-ray results are clearer than ever.

Solution by: Artificial Intelligence Institute of Innopolis University

Date: March 2021

Challenge

Chest X-ray images differ from one vendor to another. The same machine can have different configurations, and thus produce different pixel intensity distributions, making it harder to produce a generalized model. Additionally, labels may vary between datasets (Cohen, J.P., Hashir, M., Brooks, R. & Bertrand, H.. [2020]. On the limits of cross-domain generalization in automated X-ray prediction. Proceedings of the Third Conference on Medical Imaging with Deep Learning, in PMLR 121:136–155). The number of medical organizations that currently or will potentially use the system is huge, and thus it should be able to work under a heavy load.

Solution

Screening is based on a binary classification model, and visualization is based on GradCAM and saliency maps. Images are captured by X-ray machines and then transferred to the PACS of the medical organization. Then, the PACS sends anonymized images to a DICOM-adapter, which distributes data and initiates new pipelines (DAGs) for data processing. After pre-processing, data is submitted to the NVIDIA Triton inference server and then post-processed to produce valid DICOM files with visualization and a structured report. These new files are then submitted back to the PACS of a medical organization, where doctors can see the results in the tools they already use. Each medical organization can have its own set of post-processing functions to have the reports they want. The model highlights potential pathological regions in lungs.

Technologies used

• PyTorch
• NVIDIA Triton Inference Server
• Apache Airflow
• Kafka