An open-source, automated machine learning approach for large-scale image retrieval for thoracic aorta analysis studies

TL;DR

This paper introduces an open-source automated system to accurately find specific thoracic aorta CT scans from a large and varied database.

Contribution

The novel contribution is an open-source, high-accuracy image retrieval pipeline for thoracic aorta CT scans.

Findings

The pipeline identified 82% of relevant scans with 99.1% accuracy.

Excluded scans were correctly rejected in 93.6% of cases.

Abstract

To develop an image retrieval pipeline capable of identifying specific series of thoracic aortic computed tomography (CT) scans from a diverse database.

An automated image analysis pipeline was developed to select series that show the entire thoracic aorta with arterial phase contrast from within a heterogeneous institutional cohort of 4184 CT scans of the chest.

The automated pipeline identified 3435 (82%) studies from within the database that met criteria. Manual review confirmed 99.1% of the selected scans were accurately selected, and 93.6% of excluded scans were appropriately excluded.

We present an open-source, image-retrieval pipeline that, with a high degree of accuracy, can identify aortic imaging studies that meet specific criteria from within a heterogeneous collection of images. This pipeline serves as a framework that can be easily modified for other clinical use cases and can be deployed across multiple centers to promote multi-institutional research.

Figures