Nima Ashjaee

Biomedical engineer · PhD candidate, UBC

Spine models built from the images clinicians already take.

I turn routine radiographs and CT into subject-specific musculoskeletal models of the spine, estimate the loads it carries after deformity surgery, and test whether those loads explain failure. I also build the deep-learning tools that make this pipeline automatic.

Download CV Email me
Musculoskeletal model standing upright Musculoskeletal model in forward flexion Musculoskeletal model in deep forward flexion
A subject-specific thoracolumbar musculoskeletal model moving from upright standing to deep flexion, simulated in ArtiSynth.
9peer-reviewed journal articles, 4 first-author since 2025
1granted US patent
48surgical patients modeled before and after surgery
17conference presentations

01 · About

From implants to spine models

I am completing my PhD in Biomedical Engineering at the University of British Columbia, co-supervised by Dr. Thomas Oxland and Dr. Sidney Fels, with the defence planned for December 2026. My thesis, Subject-Specific Musculoskeletal Modeling in Adult Spinal Deformity, follows one question from clinical image to outcome: can models built from routine imaging estimate junctional loads that carry information about proximal junctional failure beyond alignment?

Before that, my MSc at the University of Saskatchewan combined CT segmentation, QCT-based finite element modeling, and cadaveric testing to develop an intramedullary fracture implant that is now a US patent. I now work where medical image processing, computational modeling, and machine learning meet, with the aim of building tools that spine surgeons actually use.

02 · Research

From clinical image to outcome

Each study tests one link in the chain that turns a clinical image into a spinal load, and then asks whether that load says something about the patient.

The pathway my PhD builds toward: from patient data to paired pre- and postoperative models, simulated daily activities, junctional loads, and a probability of failure.

Manuscript in preparation Adult spinal deformity

A simple load profile that predicts proximal junctional failure

Proximal junctional failure is one of the most common and costly complications of adult spinal deformity surgery, and the alignment rules used to plan these operations have had limited success in predicting it. I built paired pre- and postoperative musculoskeletal models for 48 surgical patients, 13 of whom later needed revision for PJF, and simulated eight standing and forward-bending tasks.

Two quantities at the first mobile segment above the fusion, postoperative compression and the perioperative change in anteroposterior shear, separated the patients who failed from those who did well, and did so better than the eight established clinical and radiographic frameworks tested. With the muscles switched off, the signal disappeared.

To our knowledge, this is the first study to link revision-requiring PJF to patient-specific junctional loads across daily activities and test them head to head against the frameworks used in practice. It makes muscle-dependent loading a candidate mechanical biomarker for PJF.

  • ArtiSynth
  • Biplanar radiographs
  • Pre/post-op models
  • Nested cross-validation

CMBBE 2026 Medical image processing

How much landmark error can a spine model tolerate?

From CT of ten deformity patients, I generated synthetic biplanar radiographs, moved the vertebral landmarks in four controlled ways, and compared six human raters (four researchers, two spine surgeons) who labelled the images through a web platform I built. Compression estimates held up to 6.7 mm of landmark error, but shear only to 2.2 mm. That margin sets the accuracy target for automated landmark detection.

Read the paper

European Spine Journal 2026 Machine learning

Muscle parameters without CT or MRI

Most patients have radiographs but no volumetric scan of their trunk muscles. Across 250 subject-specific spine models from the Framingham Heart Study, I trained models that infer 575 muscle force capacities and 4,008 muscle path coordinates from age, sex, height, weight, and vertebral geometry. Against anthropometric scaling, shear-force error fell from 238 N to 178 N.

Read the paperCode on GitHub

Journal of Biomechanics 2025 Musculoskeletal modeling

Which muscle properties change spinal load?

Using OpenSim, I swapped four subject-specific muscle properties for generic values, one at a time, across 250 models and 11 standing and flexed postures. Muscle path geometry and maximum isometric force changed predicted compression the most (13% and 8% on average), and the effect depended on posture. That tells us where subject-specific data is worth collecting.

Read the paper

JOR Spine 2025 Perspective

The case for subject-specific models in deformity surgery

A position paper written with spine surgeons and biomechanists from UCSF, Harvard Medical School, ETH Zurich, UC San Diego, and the University of Guelph. It grew out of the Spine meets clinic symposium I co-chaired at CMBBE 2024 in Vancouver, and proposes a two-phase pathway from understanding junctional loading to comparing surgical plans.

Read the paper (open access)

03 · Projects

Engineering and AI

AI and software

2026 · Computer vision

Automatic vertebral landmarking on biplanar radiographs

Landmark error is the weakest link in the image-to-model pipeline: shear estimates tolerate only about 2 mm. This PyTorch model finds the T1–L5 vertebral centroids in paired lateral and frontal radiographs so that nobody has to click them. Both views share a ResNet encoder and condition each other through cross-view FiLM, and DSNT heatmaps return sub-pixel coordinates with a visibility flag per vertebra.

On held-out pairs, mean radial error was 6.6 px at 384 px input (PCK@0.05 of about 0.95). DINOv2 and RAD-DINO backbones are integrated, and the model is served through a FastAPI endpoint.

  • PyTorch
  • DSNT
  • DINOv2
  • RAD-DINO
  • FastAPI

2025–present · Software

Zyra, a spine-deformity planning platform

A research workstation that goes from calibrated AP and lateral radiographs to vertebral landmarks, sagittal and spinopelvic alignment, and side-by-side construct scenarios, with load estimates from a headless ArtiSynth service and a mobile companion app.

  • Next.js
  • TypeScript
  • Expo
  • ArtiSynth

2026–present · MedARC open science

Medical foundation models

Co-author of nanopath, an open framework and benchmark for training pathology foundation models on a single H100 GPU (under review, NeurIPS 2026 workshop). For a 3D brain MRI foundation model (FOMO26 challenge), I built streaming data loading and 8×H100 distributed pretraining.

nanopath on GitHub

Implants and finite elements MSc, University of Saskatchewan, 2015–2018

Implant design · US Patent 12,082,847

An intramedullary implant for wrist fractures

Linked rods and porous spheres fill the distal radius from the inside. I iterated the design on CT-derived 3D-printed bones and in cadaveric trials, and tested fracture stability under cyclic loading on a wrist simulator I built.

View the patent

QCT-FE · Bone remodeling · ICORS 2019

How bone adapts around each implant

Subject-specific QCT-based finite element models of the radius, coupled with strain-energy bone remodeling, predicted how density would change over time with the volar locking plate and with the new implant. The implant appeared to preserve more bone, a sign of less stress shielding.

Numerical methods · CSB 2018

Faster bone-remodeling simulation

Solving remodeling at nodes is smooth but slow, and solving it at integration points is fast but leaves checkerboard artifacts. My hybrid approach matched the node-based density without the checkerboard in about one thirteenth of the run time. The same integration-point idea sped up QCT material mapping in the tibia.

Tibia mapping paper

04 · Publications

Papers and patent

Full list on Google Scholar.

  1. 2026
    Sensitivity of spinal load estimates to vertebral landmark error in biplanar radiograph–based musculoskeletal models

    Ashjaee N, Fels S, Street J, Oxland TR. Computer Methods in Biomechanics and Biomedical Engineering.

  2. 2026
    Machine learning outperforms anthropometric scaling in predicting muscle parameters and spinal loading

    Ashjaee N, Street J, Fels S, Oxland TR. European Spine Journal.

  3. 2025
    Subject-specific musculoskeletal modeling: the future of predicting and preventing proximal junctional failure in adult spinal deformity

    Ashjaee N, Semonche A, Mikula AL, Kiss L, Anderson DE, Ignasiak D, Brown SHM, Street J, Fels S, Ward SR, Ames C, Oxland TR. JOR Spine 8(4):e70142.

  4. 2025
    Effects of using generic vs. subject-specific muscle properties on spinal load prediction across different posture simulations

    Ashjaee N, Fels S, Street J, Oxland TR. Journal of Biomechanics 188:112741.

  5. 2022
    A comparative study of bone remodeling around hydroxyapatite-coated and novel radial functionally graded dental implants using finite element simulation

    Jafari B, Katoozian HR, Tahani M, Ashjaee N. Medical Engineering & Physics 102:103775.

  6. 2021
    QCT-FE modeling of the proximal tibia: effect of mapping strategy on convergence time and model accuracy

    Ashjaee N, Hosseini Kalajahi SM, Johnston JD. Medical Engineering & Physics 88:41–46.

  7. 2021
    Estimation and assessment of sagittal spinal curvature and thoracic muscle morphometry in different postures

    Pai S A, Zhang H, Ashjaee N, Street J, Wilson D, Brown SHM, Fels S, Oxland TR. Proc IMechE, Part H 235(8):883–896.

  8. 2021
    Evaluation of La(XT), a novel lanthanide compound, in an OVX rat model of osteoporosis

    Di Y, Wasan EK, Cawthray J, et al., Ashjaee N, et al. Bone Reports 14:100753.

  9. 2017
    Introduction of maximum stress parameter for the evaluation of stress shielding around orthopedic screws in the presence of bone remodeling process

    Hosseinitabatabaei S, Ashjaee N, Tahani M. Journal of Medical and Biological Engineering 37(5):703–716.

  10. Patent
    Implant for bone fracture stabilization

    Johnston GHF, Ashjaee N, Johnston JD. US Patent 12,082,847 B2, granted 2024.

  11. Review
    Nanopath: fast, fair, open experimentation for pathology foundation models

    Scotti PS, Muhammad H, Kim R, Ashjaee N, et al. Under review, NeurIPS 2026 Workshop (ASCI).

05 · Talks and service

Conferences

Teaching

Guest lecturer on finite element and multibody modeling in BMEG 330 at UBC (three offerings, first in March 2023). Teaching assistant at UBC for APSC 172, MECH 221, MECH 325, and BMEG 357 (2025–2026). Sessional lecturer for Dynamics II at the University of Saskatchewan (2019).

Service

Industry liaison for the UBC Biomedical Engineering Graduate Association (2022–2023). Program coordinator for the UBC Vancouver Summer Program (2022).

06 · Contact

Get in touch

I am looking for postdoctoral work at the intersection of medical imaging, computational modeling, and spine surgery, starting in 2027.

nima@ashjaee.com