38 - Automated Phlebolith and Fluid-Fluid Level Biomarker Detection in Venous Malformations

Monday, March 31, 2025

12:10 PM – 1:10 PM CT

Location: Music City Center, Expo Poster Area

Poster Presenter(s)

Samuel Lefcourt, MS

PhD Student
Johns Hopkins University, United States

Author/Co-author(s)

Alan J. Kim, BS (he/him/his)

Medical Student
Johns Hopkins University School of Medicine, United States
CW

Clifford R. Weiss, MD

Professor
Department of Radiology, Radiological Science and Biomedical Engineering, Johns Hopkins University School of Medicine, United States
CJ

Craig Jones, PhD

Assistant Research Professor
Johns Hopkins University, United States

Purpose: Venous malformations (VMs) are low-flow vascular anomalies which introduce pain and visual discomfort to those who have it. Treating these lesions requires an expert to meticulously leverage all available information to select the best treatment option. Phlebolith and fluid-fluid levels are key imaging findings that contribute to the characterization of a VM and can be used to differentiate low-flow vascular malformations from other vascular anomalies. By automating the detection of these biomarkers in the T2-weighted MRI volumes of head and neck patients, we provide a fast and accurate mechanism that can be leveraged within a treatment prediction model.

Materials and Methods: This is a retrospective study that used T2-WI MRI of 100 pre-treatment and 84 post-treatment scans from patients receiving percutaneous sclerotherapy for VMs of the head and neck. All VMs in the dataset were manually segmented using 3D Slicer Version 4.11. Phlebolith detection: The venous malformation was cropped from the image and rescaled to 0-255. A mask was created of all pixels with a signal intensity less than the fifth percentile. All connected components less than 7 pixels or those that had a circularity > 1 were removed. Fluid-fluid level identification: A vertical gradient magnitude image of the VM was created. For each set of three rows in the gradient image, the number of pixels in the window between the 70th and 95th percentile was counted. If the total count was greater than the number of pixels in the window, then it was identified as a fluid-fluid level.

Results: The phlebolith identification algorithm demonstrated a 75% accuracy, sensitivity and specificity when tested on a random sample of 40 patients, half of whom had phleboliths. The fluid-fluid level detector yielded a 72.5% accuracy with a sensitivity of 80% and specificity of 65%, given a separate random sample of 40 patients, half of which display fluid-fluid levels.

Conclusion: We have demonstrated early feasibility of identifying venous malformation imaging biomarkers automatically and efficiently. Automated phlebolith and fluid-fluid level detection has the potential to facilitate imaging-based VM diagnosis and will benefit future work towards developing of an imaging-based predictive model for VM outcomes.