A DICOM Framework for Machine Learning and Processing Pipelines Against Real-time Radiology Images.

Area of Science:

• Medical Imaging
• Machine Learning
• Health Informatics

Background:

• Executing machine learning (ML) pipelines on radiology images is challenging due to limited clinical computing resources.
• Research clusters require efficient data transfer from hospital Picture Archiving and Communication Systems (PACS).

Purpose of the Study:

• To introduce Niffler, an open-source Digital Imaging and Communications in Medicine (DICOM) framework.
• To enable real-time ML and processing pipelines in research clusters by efficiently retrieving images and metadata from PACS.

Main Methods:

• Developed Niffler, an open-source DICOM framework for data retrieval and metadata extraction.
• Deployed Niffler at Emory Healthcare, retrieving up to 350 GB/day of DICOM data over two years.
• Utilized Niffler for real-time ML pipeline execution, bulk on-demand image retrieval, and metadata analysis.

Main Results:

• Successfully executed a real-time IVC filter detection and segmentation pipeline on abdominal radiographs with 96.0% accuracy.
• Validated Niffler's accuracy in identifying MRI scanner examination timeframes and idling times against Clinical Data Warehouse (CDW).
• Identified and corrected misconfigured scanner times on five systems using Niffler-extracted metadata.

Conclusions:

• Niffler effectively enables real-time machine learning and processing pipelines in research clusters.
• The framework enhances data retrieval efficiency and metadata extraction from DICOM images.
• Niffler demonstrates significant utility in clinical research, operational efficiency analysis, and system quality control.