Skip to main content
2024 BMES Annual Meeting Main banner
Icon Legend
This session is not in your schedule.
This session is in your schedule. Click again to remove it.

Bioinformatics, Computational and Systems Biology

Bioinformatics, Computational and Systems Biology - Poster Session A

Poster A8 - Automating an Image Processing Pipeline to Classify and Segment HeartShare CT Scans

Thursday, October 24, 2024
10:00 AM - 11:00 AM EST
Location: Exhibit Hall E, F & G

Introduction:  HeartShare is an enterprise funded by the National Heart, Lung, and Blood Institute (NHLBI) that performs extensive analysis of phenotypic medical images, data, and omics collected from a multi-center cohort of individuals diagnosed with heart failure with preserved ejection fraction (HFpEF) and age and sex-matched controls. HeartShare participants receive chest computed tomography (CT) scans, which can be used to quantify characteristic features of the lungs. CT scans are collected with the same imaging protocol by 8 different clinical sites, but across sites there is inconsistent categorization and storage of the CT scan series’, leading to increased risk of user error and increasing time needed to sort and process the scans for lung outcomes. The heterogeneity of the CT data stored as Digital Imaging and Communications in Medicine (DICOM) files traditionally requires inefficient manual categorization and segmentation of these scans, patient by patient. Addressing these challenges, this study aimed to automate the image processing pipeline to enhance efficiency and precision of the sorting, organization, and segmentation.

Materials and

Methods:  Imaging data was obtained from the HeartShare program. Python (v3.9.0) and MicroDICOM (v2022.1) were utilized to extract image components from the header (or image metadata) such as the slice thickness, series description, and convolution kernel. Extracted features became differentiators for classification using a recursive detection scheme. Figure 1 shows the unique directory structures used to store image series across four different clinical sites out of eight. Each directory used can have different folder names, lengths, and naming conventions which emphasize the need for a streamlined output. The pipeline gathers the DICOM images from these scans, interprets the key image components, and produces an easily interpretable and uniform format.
Following pre-processing, 3D Slicer (v5.6.2) was employed to perform semi-automated segmentation of the lungs using the Chest Imaging Platform (CIP) [2][3]. To extract the desired metrics from CIP, one image volume was loaded into Slicer and the CIP automated segmentation component and statistics analysis were started manually. Figure 2 demonstrates the CIP-produced segmentation of a lung scan. In contrast, fully automated, batch processing was used to segment the lungs and extract physical characteristics using an in-house python script for TotalSegmentator [4]. Comparison of lung volumes and the processing time for 181 subject CT scans (Age: 68 ± 13 years; BMI: 31 ± 7 kg/m2; 59.7% female; 82.87% white) processed with CIP vs. TotalSegmentator were performed to evaluate efficacy of both image processing methods. A paired t-test was used to examine the lung volume differences in a subset of HeartShare patients.

Results, Conclusions, and Discussions: The implementation of the DICOM categorization script yielded an efficient and accurate set of patient directories which can be used in both a manually assisted, semi-automated or fully manual segmentation system. The semi-automated CIP takes ~20 minutes/scan to process, including transferring, loading and processing/refinement of each scan, while the fully automated TotalSegmentator took ~ 2.5 minutes/scan, resulting in an 87.5% percent increase in efficiency when utilizing the automated process. For 181 scans, the automated TotalSegmentator pipeline resulted in over 52 hours of time saved.

Furthermore, for the segmentation, Table 1 shows the results of this test. Both methods produced similar lung volumes (5.07 ± 1.21 L vs 4.96 ± 1.25 L), with small percent differences (2.29 ± 7.80%). While a paired t-test found statistically significant volume differences between the two methods (2.29 ± 7.80%, p=0.0003), the difference was not clinically meaningful as it was deemed to be within the resolution limits of the CT scans.

The results from this study suggest an encouraging basis for the automation of these categorization and segmentation tasks. The pre-processing application was applied to this research and will benefit future work as the HeartShare cohort continues to grow. Moving forward, further optimization and validation would be essential to enhance the script’s performance. Nevertheless, the amount of time preserved through this method proved to be both worthwhile and efficacious.

Acknowledgements (Optional):  This work was supported by the NHLBI HeartShare Program through the following grants: U54 HL160273 (Northwestern University Data Translation Center); U01 HL160279 (Northwestern University); U01 HL160277 (University of Pennsylvania); U01 HL160274 (University of California at Davis); U01 HL160226 (Mayo Clinic); U01 HL160272 (Wake Forest); U01 HL160278 (Massachusetts General Hospital) and by the FNIH Accelerating Medicines Partnership Heart Failure (AMP HF) Program [including RFP 2023-1345-001 (Johns Hopkins University). This project was supported in part by the NIH/NLM (R25 LM014214) in the Department of Biomedical Engineering and Center for Artificial Intelligence Research at Wake Forest University School of Medicine.