Research · Submitted to Radiology: Cardiothoracic Imaging
Tear Detector
Finding flow-active communications in aortic dissection—and extracting their size and time course from 4D Flow MRI.
My contribution
Locating and measuring flow-active tear candidates.
I developed the detector to connect candidate selection, size estimation, and a time-resolved signal after lumen segmentation, with evaluation in a clinical research cohort.
Lead inventor of Fen-Map. A provisional patent application has been filed. Northwestern’s technology transfer office is pursuing a nonprovisional application.
Clinical problem
Finding and measuring active tear communications.
In aortic dissection, tears in the intimal flap connect the true and false lumens. Anatomical imaging depicts these openings; 4D Flow MRI adds how blood moves between them over time.
The measurement problem
Conventional tear-flow analysis begins with an observer recognizing an opening, choosing a communication, and positioning a measurement plane. A flow-guided search starts from the velocity field to identify candidate sites for characterization.
Why the data are difficult
Three velocity components vary across a curved anatomy and multiple cardiac frames. Their meaning depends on local lumen geometry. Spatial resolution, segmentation, and the strength of the flow signal all influence which voxels are selected.
Computational method
From a velocity field to a candidate-level measurement.
The contribution is an integrated, flow-guided workflow: search the lumen boundary, select a candidate, then derive its extent and time course from the same voxels.
From lumen anatomy to selected tear voxels

The cyan region marks the false-lumen side of the computational boundary. Flow colorbars show voxelwise peak signal in mL/frame; numbered shapes illustrate clusters rather than quantitative rankings.
Open full-size figure (new tab)Define the anatomy
Lumen masks + corrected velocities
True- and false-lumen segmentations supply the geometry. Centerlines establish a local direction between the two lumens.
Isolate cross-lumen flow
A direction-aware flow map
Project the velocity field toward the neighboring lumen, perpendicular to the false-lumen centerline. This emphasizes cross-lumen motion.
Find candidate clusters
The computational lumen boundary
Search the false-lumen side of the shared boundary. Threshold peak true-to-false-lumen signal, group adjacent voxels, and remove small clusters.
Measure the selected site
One cluster for all measurements
Select a cluster using flow magnitude and extent. Estimate its long-axis length and sum the same voxels at every recorded frame.
Technical definition and selection rule
The projection direction points from the false-lumen centerline toward its nearest true-lumen centerline, with the false-lumen tangent component removed. It is derived from centerlines, rather than a measured local flap normal. Corrected velocities are interpolated to a 1-mm grid; voxel-face area and frame duration convert the projection to a voxelwise signal in mL/frame.
The detector thresholds each boundary voxel’s largest true-to-false-lumen magnitude across frames. Clusters are ranked by mean peak magnitude; the longer of the leading two is selected. Principal-component analysis defines its long axis, measured between projected outermost voxel centers. A selected “dominant” cluster reflects this rule, not established physiological importance.
Output structure
A location, a size estimate, and a flow signal.
All three outputs refer to the same selected boundary voxels, connecting the spatial measurement to its time course.
Selected candidate
One cluster carried through each measurement.
- Spatial support
- The selected voxel coordinates identify where the candidate lies on the computational boundary between the true and false lumens.
- Long-axis extent
- The distance between outermost voxel centers along the cluster’s principal axis provides a three-dimensional size estimate.
- Time-resolved signal
- Summing the signed signal over those same voxels at each recorded frame describes how the selected candidate’s signal changes over time.
The time course has not been independently validated as total flow through the tear. The cluster length is a flow-derived size estimate, and anatomical correspondence still requires evaluation.
Evidence and limits
Internal evaluation in 54 patients.
The current study compares automated estimates with manual angiographic tear-size measurements at the patient level. The measurements were not confirmed to represent the same anatomical tear.
- Study cohort
- 54 patients
- Retrospective, single-center study of descending aortic dissection.
- Output availability
- 50 of 54
- Patients with a selected cluster and time course. This is not a measure of detection sensitivity.
- Mean absolute size difference
- 4.22 mm
- Leave-one-patient-out evaluation; 95% confidence interval, 3.20–5.34 mm.
Each held-out patient’s detector settings were fitted using the other 53 patients. All patients, including recorded zeros, contributed to the size comparison. Earlier preprocessing and detector design choices were not reevaluated within these folds, so this remains internal evaluation of a developed method.
Patient-level size comparison

Leave-one-patient-out estimates, including recorded zeros. The right panel shows signed differences and descriptive agreement limits; shaded bands are patient-bootstrap 95% confidence intervals. These are not confirmed measurements of corresponding tears.
Open full-size figure (new tab)What remains unresolved
Flow-guided selection can identify a different communication from anatomical review. Substantial patient-level differences mean the size estimates should not be treated as interchangeable with manual measurements. The current study does not establish detection accuracy or clinical utility.
Validation still needed
Matched anatomical tear localization, independent flow measurements, external cohorts, segmentation repeatability, and scan-rescan testing are needed to define when these outputs are reliable.
Future direction
Structured measurements for a reviewable analysis system.
A candidate’s location, extent, and time course could become inputs to a future network of analysis agents, alongside measures of uncertainty.
Specialized agents could check input quality, inspect candidate consistency, compare measurements across time, and assemble an evidence summary for an investigator. Each conclusion would need to remain traceable to its source images and measurements, with uncertain cases flagged for review.
An initial experiment could test whether an agent flags inconsistent candidate measurements and points a reviewer back to the relevant images. Its findings would be compared with expert review; this workflow has not yet been implemented or evaluated.
Companion research
Making the dissection flap easier to inspect.
Flap-specific curved planar reformatting addresses a complementary problem: presenting complex dissection anatomy for visual review.
Explore curved planar reformattingSubmitted to Radiology: Cardiothoracic Imaging
Automated 4D Flow MRI-Based Dominant Entry Tear Detection and Risk Mapping in Descending Aortic Dissection
Cohn S, Weiss EK, Mehta CK, Markl M, Allen BD.