Christian
Rogers

Coronary CT, measured at the resolution it was acquired.

Segmentation tooling and deep learning for coronary imaging at the So Lab, Lawson Research Institute and Western University. Low level runtime work on the side.

Current work GitHub LinkedIn

Ongoing

Asclepius

Multiclass coronary artery segmentation from CCTA, with no downsampling anywhere in the path.

A distal branch is 1.5 to 2 mm across, a few voxels at the 0.35 mm the scans are acquired at. Anything that coarsens the grid removes the structure being labelled, so the model reads the data at acquired resolution end to end. No cascade, no heart crop, no mirroring.

Follow the work

Model: nnU-Net v2, 3d_fullres, native spacing
Compute: SciNet Trillium, one H100 80 GB per job
Patch: About 256 cubed, near 30% of a volume
Stage one: Binary lumen on 1000 ImageCAS masks
Benchmark: 82.96% Dice, published

Open to contributors

Two ways in. Segmentation labelling, marking per branch coronary anatomy, and engineering, on the training and evaluation pipeline.

christian.rogers@mail.utoronto.ca

Publication

Fractional Flow Reserve from Dynamic CT Perfusion, in Heavy Calcification and Stents

So A, Nakamura S, Takafuji M, Haider M, Rogers C, Teefy P, Kitagawa K.
Diagnostics 16(15), 2408. MDPI, 2026.

Six panel figure: a straightened vessel view marked with pre lesion, lesion entry, lesion exit and post lesion points; a three dimensional coronary tree with a stented segment; two axial CT slices at the stenosis; and two time enhancement curves in Hounsfield units. — Lesion points on a straightened vessel view, the coronary tree with a stented segment, and the time enhancement curves the FFR is computed from. Figure 2, CC BY 4.0.

A CT dynamic angiographic imaging algorithm derives FFR from stress myocardial perfusion, in the vessels that calcium and prior stents usually rule out. Values were compared against invasive catheter FFR in 16 patients.

0.974Correlation R

100%Sensitivity

100%Specificity

0.004Mean difference

Calcium score of the stenoses: 1911.63 Agatston Units. Bland Altman limits of agreement: -0.061 to 0.069.

Read the paper

Selected work

DAI Workflow for 3D Slicer

Python / Research

Guided segmentation that extracts lesion points and vessel radius. Cropping, vessel segmentation, centerline extraction and curved planar reconstruction in one path, built for clinical pace and minimal training.

3D SlicerCenterline extractionCPRSo Lab

globals

TypeScript / npm

Shared application state across Electron processes on one region of real shared memory. Reads are synchronous with no await and no round trip, writes are asynchronous, the lock is a non blocking seqlock. Prebuilt for macOS, Linux and Windows.

ElectronShared memoryNative addonMIT

Elsewhere

Emailchristian.rogers@mail.utoronto.ca GitHubchristianGRogers LinkedInChristian Rogers Diagnostics 202610.3390 / diagnostics16152408