Back to Research

Thesis I · Sacral MRI · Sep 2025 – Mar 2026

Automatic lumbosacral vertebra localization in variable field-of-view pelvic MRI

A Deep Learning Approach for Variable Field-of-View Pelvic Imaging

Author: Saad AhmadInstitution: FAU Erlangen-NurembergLab: Smart Imaging LabSupervisor: Prof. Dr. Jana Hutter

4.21 mm

Mean Localization Error

257 landmarks, multi-center

3.77 mm

S1 Anchor Error

100% within 10 mm

3.54 mm

Median Error

Across all protocols

222

Training Cases

Protocol II + III

The Problem This Work Solves

Accurate lumbosacral vertebrae localization in pelvic MRI is clinically critical for spinal surgery planning, radiation therapy field definition, endometriosis staging, and large-scale morphometric studies. Yet existing automated methods fundamentally fail in this setting — they assume full spinal visibility from C2 or T1, which is never available in pelvic MRI.

In pelvic imaging, the spine appears only incidentally: 84.7% of cases show partial or sacral-only coverage. No prior work had addressed individual S1–S5 sacral vertebra detection in this context, leaving a critical gap in automated pelvic MRI analysis.

The Innovation: S1-Anchored Detection

The core insight is that pelvic MRI always captures the S1 vertebra — it is the anatomical bridge between the sacrum and the pelvis. Rather than requiring a superior reference (the conventional approach), this work proposes making S1 the anchor point for bidirectional vertebral labeling, implemented through a novel two-stage system:

  • Dual-Head 3D U-Net: A lightweight architecture (~400K parameters) with a shared encoder-decoder and two task-specific output heads — one detecting all vertebral centers, one dedicated to S1 morphology. Joint training enables the model to learn both global vertebral patterns and S1-specific anatomy simultaneously.

  • S1-Anchored Post-Processing: A multi-stage inference pipeline that detects S1 from its dedicated heatmap, then propagates labels bidirectionally — cranially through lumbar vertebrae and caudally through sacral segments — without any need for upper vertebrae to be present in the scan.

  • Biased Patch Sampling: 70% of training patches are drawn from high-activity heatmap regions, forcing the model to learn from challenging anatomical boundaries rather than uniform sampling.

Architecture

Inference Pipeline

Pelvic MRI (NIfTI)Isotropic ResamplingPatch Extraction (biased)Dual-Head 3D U-NetS1 Heatmap → AnchorBidirectional Label PropagationL4–S5 Landmarks

Dual-Head 3D U-Net

Input Volume — 3D MRI Patch

Shared Encoder

Conv3D + BN + ReLUResidual Block ×2Max Pool ↓2Residual Block ×2Max Pool ↓4Bottleneck

Head 1 · All Vertebrae

Decoder + Skip Connections
ConvTranspose3D ×3
1×1×1 Conv
Multi-class Heatmap (L4–S5)

Head 2 · S1 Dedicated

Decoder + Skip Connections
ConvTranspose3D ×3
1×1×1 Conv
S1 Anchor Heatmap

S1-Anchored Post-Processing

Peak Extraction (S1)Bidirectional PropagationCaudal: S2→S5Cranial: L5→L4
Output — L4, L5, S1, S2, S3, S4, S5 Landmark Coordinates
~400K ParametersAdam OptimizerReduceLROnPlateauMSE Heatmap LossGaussian σ=2mmBiased Patch Sampling 70%

Per-Vertebra Results

VertebraMean Error≤5 mm≤10 mmBar
L44.36 mm70.0%90.0%
L53.65 mm78.9%97.4%
S1 (Anchor)3.77 mm75.0%100%
S24.33 mm75.0%97.7%
S33.54 mm83.7%97.7%
S44.79 mm69.2%89.7%
S55.25 mm62.5%87.5%

S1 anchor achieved 100% detection within 10 mm across all 44 test instances. S4–S5 show higher variance due to progressive developmental fusion.

Contextualized Against Prior Work

MethodSettingPerformance
Existing methodsRequire full spinal visibility (C2/T1 reference)Not applicable
Glocker et al. (CT)Full-spine CT, regression forests8.6 mm
Windsor et al. (MRI)Spine-dedicated MRI, full FOV2–4 mm (lumbar only)
This workPelvic MRI, partial FOV, all sacral levels4.21 mm (L1–S5, multi-center)

Results are within the range of spine-dedicated methods despite operating on a fundamentally harder problem: partial spinal coverage, lower through-plane resolution, and variable patient positioning.

Clinical Impact & Future Directions

This is the first published method for individual S1–S5 sacral vertebra detection in pelvic MRI, enabling:

  • Reproducible lumbosacral morphometry in gynecological imaging workflows — reducing inter-reader variability in surgical and radiotherapy planning.

  • Observer-independent spinal measurements at population scale unlocking large-scale epidemiological studies previously blocked by manual bottlenecks.

  • A modular building block for joint pelvic analysis systems that co-localize vertebral and uterine anatomy from a single MRI acquisition.

Future work is planned toward a unified pelvic MRI framework that jointly predicts both spinal and uterine landmarks, leveraging the geometric relationship between the sacrum and uterus to mutually constrain predictions.

PythonPyTorch3D U-NetNIfTI / nibabelGaussian Heatmap RegressionSciPy3D SlicerAdam + ReduceLROnPlateauMulti-center evaluation