Automated Landmark Detection for assessing hip conditions: A Cross-Modality Validation of MRI versus X-ray

Roberto Di Via, Vito Paolo Pastore, Francesca Odone, Siôn Glyn-Jones, Irina Voiculescu

TL;DR

This paper presents the first cross-modality validation study comparing automated deep learning-based landmark detection on paired X-ray and MRI scans for FemoroAcetabular Impingement (FAI) screening. The authors demonstrate that standard heatmap regression models trained on MRI achieve equivalent localization and diagnostic accuracy to traditional X-ray approaches, establishing a validated 2D foundation for future 3D volumetric hip analysis.

What Problem Does This Paper Address?

FemoroAcetabular Impingement (FAI) is a painful hip disorder affecting 20–25% of the population, where abnormal contact between the femoral head and the acetabular rim leads to cartilage damage and early osteoarthritis. Clinicians diagnose FAI using geometric angles calculated from specific anatomical landmarks: the $\alpha$-angle (measuring femoral head asphericity, or cam morphology) and the lateral centre-edge (LCE) angle (measuring acetabular coverage, or pincer morphology).

Traditionally, these angles are measured on 2D anteroposterior (AP) pelvic X-rays. However, X-rays suffer from projection limitations; variations in pelvic tilt or patient alignment can alter angle measurements by up to $15^\circ$, potentially misclassifying borderline cases. While 3D MRI scans offer a clearer, correctable view of the anatomy, automated landmark detection pipelines have primarily been validated only on X-rays. This paper addresses this gap by investigating whether geometric FAI metrics can be reliably and equivalently extracted directly from MRI.

Methods

The authors frame FAI angle computation as a landmark detection problem, targeting four key anatomical keypoints per hip:

  1. Femoral head centre (FHC)
  2. Neck-axis point (NA)
  3. Lateral acetabular edge (LAE)
  4. Lateral cam point (LCP)
Spatial correspondence of annotated landmarks across modalities

They utilize a standard heatmap regression architecture, converting the coordinate detection task into predicting ground-truth Gaussian heatmaps ($\sigma=5$). The core model evaluated is a UNet++ architecture with an ImageNet-initialised ResNet18 encoder, trained using Negative Log-Likelihood (NLL) loss. To boost robustness during inference, they implement Test-Time Augmentation (TTA) by averaging heatmap predictions across multiple stochastically transformed views of the test images.

Datasets

The study utilizes a paired dataset of 89 patients collected from the Femoroacetabular Impingement Trial (FAIT). Each patient underwent both an AP pelvic X-ray and a multi-slice hip MRI.

For the MRI data, the authors focus on T1-weighted coronal acquisitions (20 slices with 3.3 mm spacing). To establish a direct 2D comparison with X-rays, all annotations were standardized to the middle slice (slice index 10), which represents the mid-acetabular coronal plane where all four landmarks are concurrently visible. The dataset was split at the patient level into training, validation, and testing sets (105:17:40 images, representing a 65:10:25 split).

Key Insights and Ideas

  • Complementary Modality Strengths: The two imaging modalities exhibit distinct, complementary advantages based on tissue contrast. X-rays excel at localizing the femoral head centre (FHC) due to high bone-air contrast, whereas MRI yields superior accuracy for the neck-axis (NA) point due to its enhanced soft-tissue delineation.
  • The Elusive Cam Point: The lateral cam point (LCP) remains highly challenging to pinpoint for both deep learning models and human clinicians alike. This inherent ambiguity results in higher mean errors across both modalities, though median errors remain low, indicating that the mean is skewed by a few borderline outliers.
  • Clinical Equivalence: Despite localized differences in landmark placement errors, both modalities ultimately converge on the exact same diagnostic accuracy for cam-type impingement screening.

Key Results

  • Landmark Localisation: MRI achieved an overall Mean Radial Error (MRE) of 2.98 ± 0.23 mm, closely matching the X-ray MRE of 3.02 ± 0.10 mm.
  • Diagnostic Accuracy: For cam-type impingement screening ($\alpha > 65^\circ$), both X-ray and MRI models achieved an identical diagnostic accuracy of 87.50%, with perfect specificity (100.00%) and identical sensitivity (54.55%).
  • Angle Agreement:
  • For the LCE-angle, MRI demonstrated a lower median error ($1.28^\circ$ vs $2.22^\circ$ for X-ray) and excellent clinical reliability (Intraclass Correlation Coefficient, ICC = 0.73).
  • For the notoriously difficult $\alpha$-angle, both modalities achieved fair clinical significance (X-ray ICC = 0.52, MRI ICC = 0.41), mirroring the typical inter-observer variability seen among human clinicians.
FAI assessment from paired T1 MRI and X-ray

Limitations and Future Work

The primary limitation of the current work is its reliance on a single 2D MRI slice (the mid-acetabular plane), which underutilizes the rich volumetric data inherently provided by a 3D MRI scan and ignores inter-slice coherence. Additionally, the cohort size (89 patients) is relatively modest due to the clinical difficulty of obtaining longitudinal, multi-modality paired datasets.

Future work aims to extend this framework to true 3D volumetric analysis, computing metrics averaged across multiple slices or directly from 3D volumes. The authors also plan to develop automated slice selection methods to eliminate the manual standardisation step currently required.

Summary updated

Abstract

Many clinical screening decisions are based on angle measurements. In particular, FemoroAcetabular Impingement (FAI) screening relies on angles traditionally measured on X-rays. However, assessing the height and span of the impingement area requires also a 3D view through an MRI scan. The two modalities inform the surgeon on different aspects of the condition. In this work, we conduct a matched-cohort validation study (89 patients, paired MRI/X-ray) using standard heatmap regression architectures to assess cross-modality clinical equivalence. Seen that landmark detection has been proven effective on X-rays, we show that MRI also achieves equivalent localisation and diagnostic accuracy for cam-type impingement. Our method demonstrates clinical feasibility for FAI assessment in coronal views of 3D MRI volumes, opening the possibility for volumetric analysis through placing further landmarks. These results support integrating automated FAI assessment into routine MRI workflows. Code is released at https://github.com/Malga-Vision/Landmarks-Hip-Conditions
BibTeX
@inproceedings{via2026automated,
  author = {Roberto Di Via and Vito Paolo Pastore and Francesca Odone and Siôn Glyn-Jones and Irina Voiculescu},
  title = {{Automated Landmark Detection for assessing hip conditions: A Cross-Modality Validation of MRI versus X-ray}},
  booktitle = {ISBI},
  year = {2026},
  doi = {10.1109/isbi61048.2026.11515938},
}
Figure