Thesis II · Pelvic MRI · Sep 2025 – Mar 2026
Automatic Detection of Uterine
Anatomical Landmarks in Pelvic MRI
A Segmentation-Guided Multi-Decoder 3D U-Net with SharpHeatmapLoss
4.52 mm
Overall Mean Error
192 landmarks evaluated
2.90 mm
Best: Cavity Cervix
sub-3 mm precision
3.66 mm
Median Error
All 6 landmarks
32
Test Cases
3 acquisition protocols
Problem Statement
Uterine biometry — measuring fundal thickness, body length, cervical length, and AP diameter — is essential for managing endometrial cancer, fibroids, and endometriosis. Yet it is performed manually with high inter-observer variability. No prior automated 3D approach existed for multi-landmark localization from a single pelvic MRI acquisition.
The six anatomical targets — APD-1, APD-2, Fundus Outer, Cavity Cervix, Inner OS, and Cavity Fundus — span the full uterine extent and vary significantly in appearance across acquisition protocols, making a single-model approach particularly challenging.
Approach
Proposed a segmentation-guided multi-decoder 3D U-Net predicting all six uterine landmarks simultaneously via dedicated decoder branches. Three key innovations drive performance:
Geometrically Consistent Augmentation: In-plane rotations and left-right flips with coordinate transformations across LPS, RAS, and voxel space ensure landmark annotations remain valid after augmentation, exposing the model to the full range of uterine tilt and orientation seen in real patients.
ROI-Guided Inference: Uterine segmentation masks are used to crop the inference region, eliminating background noise and reducing false activations from pelvic structures. The model only processes voxels within the uterine envelope.
Dual-Model Architecture: Two custom 3D U-Nets with shared encoders for contextual learning of landmark relationships, and independent decoder branches per landmark group: Model 1 for APD-1, APD-2, and Fundus Outer; Model 2 for Cavity Cervix, Inner OS, and Cavity Fundus. The shared encoders capture inter-landmark spatial context while separate decoders specialize on landmark-specific features, preventing cross-task gradient interference.
SharpHeatmapLoss: A custom loss function combining MSE with a cubic penalty term that forces sharply peaked Gaussian heatmaps. Standard MSE allows diffuse activations; the cubic term penalizes spread disproportionately, compelling the network to localize with sub-voxel precision.
Architecture
Inference Pipeline
SharpHeatmapLoss
The cubic term penalizes diffuse activations disproportionately, forcing the network to concentrate probability mass at the true landmark center.
Dual-Model Architecture
Model 1 — TwoHead U-Net
APD-1, APD-2
Fundus Outer
Model 2 — ThreeHead U-Net
Cav. Cervix
Inner OS
Cav. Fundus
Per-Landmark Results
| Landmark | Mean Error | Model | Bar |
|---|---|---|---|
| Cavity Cervix | 2.90 mm | ThreeHead | |
| Cavity Fundus | 2.94 mm | ThreeHead | |
| Inner OS | 4.85 mm | ThreeHead | |
| Fundus Outer | 5.11 mm | TwoHead | |
| APD-1 | 5.66 mm | TwoHead | |
| APD-2 | 5.66 mm | TwoHead |
Scale max ≈ 6.7 mm · Lower = better · Cavity Cervix & Fundus achieve sub-3 mm precision.
Clinical Impact & Future Directions
This is the first automated 3D approach for uterine multi-landmark localization from pelvic MRI, enabling:
Reproducible uterine biometry at scale — reducing inter-reader variability in fibroid monitoring, endometriosis staging, and surgical planning.
Sub-3 mm precision on two landmarks — Cavity Cervix (2.90 mm) and Cavity Fundus (2.94 mm), opening the door to automated cervical length measurement.
Multi-protocol generalization — evaluated across 3 acquisition protocols and 32 test cases, demonstrating robustness to scanner and sequence variation.
Future directions include joint spinal-uterine landmark prediction from a single pelvic MRI acquisition, leveraging the geometric relationship between the sacrum and uterus established in Thesis I.