1355471 - Fully Automated Data Extraction, Templating, and Visualization using Custom Algorithms for Analysis of Oxygenation-Sensitive CMR

Oxygenation-Sensitive Cardiovascular Magnetic Resonance (OS-CMR) has emerged as a powerful tool to investigate the pathophysiology of several disease states by assessing dynamic changes in myocardial oxygenation¹. Recently, the analysis of CMR scans with radiomics algorithms has demonstrated superior diagnostic accuracy over standard analytical and reporting methods². However, one of the main limiting factors in this analysis is the labour-intensive extraction and structuring of the extreme amount of raw data into a desirable format. Previous attempts to create a universal extraction package have been partially unsuccessful due to many variations in the type of data within patients and studies. An automated, personalized algorithm that could be generalized to diverse kinds of studies would allow for more efficient data extraction, structuring, visualization, and subsequent analysis.

Methods:

Two OS-CMR data sets were utilized in the development and validation of the data extraction and visualization tool. The algorithm was developed and first tested on a data set of 51 patients with suspected coronary artery stenosis and 27 healthy volunteers. The algorithm was then validated in a data set of 33 patients with ischemia with no obstructive coronary artery stenosis, and 32 healthy volunteers. All CMR data were collected on a clinical 3T Magnetom Skyra™ (Siemens Healthineers, Erlangen, Germany). A python library was created to extract and export raw patient data into a pre-established template structure compatible with most modern data analytics frameworks. The python library uses the openpyxl package to read and write on excel files and Plotly to generate the graphs. The updated template contains over 3,300+ discrete data points (compared to 602 from previous methods) per participant separated into independent OS-CMR biomarkers. The algorithm's accuracy was assessed by contrasting its output to manually extracted data on a subset (690 biomarkers) of the overall data set.

Results:

In comparison to the manually extracted data, the algorithm achieved 100% accuracy in extracting the data into the appropriate format. The algorithm's capability was also evaluated in different patient populations and proved to be successful and generalizable across disease states (Figure). Additionally, there was a significant reduction in data extraction time between the manual method (2.3 hours per participant and 179 hours for the complete data set) and automated algorithm extraction (~1 second for the entire data set).

Conclusion:

We have developed a fully automated tool for raw data extraction, structuring, and visualization that can be used to aid in big data analysis of OS-CMR studies. The potential to significantly streamline this process will allow for the application of OS-CMR in routine clinical settings.