Measurements Automatically Extracted from Zero Echo Time MRI Using Deep Learning Image Segmentation and Geometric Modeling Agree with Expert Manual Readings

Measurements Automatically Extracted from Zero Echo Time MRI Using Deep Learning Image Segmentation

arXiv:2608.07368v1 [physics.med-ph] 7 Aug 2026

and Geometric Modeling Agree with Expert Manual Readings Jack Consolini∗1 , Eric A. Bogner1 , Meghan Sahr1 , Matthew F. Koff1 , Kevin M. Koch1 , and Hollis G. Potter1 1

Department of Radiology and Imaging, Hospital for Special Surgery, New York, NY, USA

∗

Address correspondence to: Jack Consolini, M.S., Department of Radiology and Imaging, Hospital for

Special Surgery, New York, NY 10021, USA. Email: consolinij@hss.edu.

1

Abstract Background: Computed tomography (CT) remains the reference for three-dimensional osseous morphometry for patients with femoroacetabular impingement (FAI) but entails ionizing radiation and manual measurement. Zero echo time (ZTE) magnetic resonance imaging (MRI) permits visualization of cortical bone, and FAI angles from ZTE MRI agree with CT; however, automated angle extraction remains limited. Purpose: To develop and validate automated FAI angle computation from ZTE MRI that will have good to excellent agreement to manual measurements from expert musculoskeletal radiologists. Study Design: Cross-Sectional study; Level of evidence, 3. Methods: Pelvic ZTE MRI was acquired on 73 participants (mean ± standard deviation age 36.8 ± 18.5 years; 51 women/22 men) yielding 135 hips for inclusion. An nnU-Net model was trained on 100 manually curated hips to automatically segment the femur, pelvis, and three osseous landmarks. Custom-developed automated geometric analysis algorithms computed alpha, femoral neck-shaft, Tönnis, coronal and sagittal center-edge, and acetabular version angles from inferred segmentations. Measurements on 35 test hips were compared with the mean of manual measurements made by two musculoskeletal radiologists using intraclass correlation (ICC) and Bland–Altman analysis. Results: Dice coefficient exceeded 0.96 for bone segmentation. Landmark segmentation Dice accuracy ranged from 0.65–0.83. Median landmark error was 0.38 mm (femoral head), 0.82 mm (lateral acetabulum), and <2.5 mm (medial acetabulum and greater trochanter). ICC was excellent for acetabular versions, coronal center-edge, and Tönnis (ICC ≥ 0.82) but poor for alpha and femoral neck-shaft. Model versus rater-mean agreement was excellent for acetabular version, coronal center-edge, and Tönnis (0.92–0.96), good for mid-acetabular sagittal center-edge (0.74), and fair for alpha (0.45) and femoral neck-shaft (0.55). Bland–Altman limits of agreement for the model were narrower than interrater limits for most angles.

2

Conclusion: Fully automated quantitative morphometric assessment from ZTE MRI is feasible and performs comparably to an expert reader for most coverage and version angles. Clinical Relevance: This approach may reduce adjunct CT for preoperative morphometric assessment of FAI and hip dysplasia in athletes, young active patients, and women of reproductive age, providing standardized, automated angle measurements from a single radiation-free MRI examination. Key Terms: femoroacetabular impingement; hip dysplasia; magnetic resonance imaging; automated morphometry; hip preservation; deep-learning

Abbreviations AIR

Adaptive image receive

CT

Computed tomography

DICOM

Digital imaging and communications in medicine

FAI

Femoroacetabular impingement

HIPAA

Health insurance portability and accountability act

ICC

Intraclass correlation coefficient

MRI

Magnetic resonance imaging

NIfTI

Neuroimaging informatics technology initiative

nnU-Net

No-new convolutional network for biomedical image segmentation

PACS

Picture archiving and communication system

RANSAC

Random Sample Consensus

SDdiff

Standard deviation of differences

YOE

Years of experience

ZTE

Zero echo time

3

1

Introduction

Femoroacetabular impingement (FAI) and dysplasia of the hip disrupt normal joint biomechanics and are established contributors to early osteoarthritis in young and active patients 1,2,3,4,5,6,7,8 .

Abnormal femoral and/or acetabular morphology produces pathologic

contact during hip motion 1,2,3 . Common sub-types include pincer (acetabular overcoverage), cam (femoral head–neck asphericity), or combined types 9,10,11 . Hip dysplasia is characterized by insufficient acetabular coverage of the femoral head, causing edge loading of the shallow socket 7,8 and is often accompanied by a cam deformity, compounding symptoms 12,13 . Abnormal acetabular shape is common in young populations, with radiographic evidence of FAI-related morphology in 75–95% of athletes 11,14,15 and 38% of hips evaluated for pain 3 . Progressive impingement contributes to labral and chondral injury and is implicated in early osteoarthritis 16,17 , making accurate morphologic characterization essential to hip-preservation surgical decision-making 1,6,9 . Pre-operative assessment relies on radiography for screening and CT for detailed threedimensional osseous morphometry, while MRI remains essential for soft-tissue evaluation of the labrum and articular cartilage 1,10,18,19,20 . Angular measurements, including alpha, femoral neck-shaft, acetabular version, and center-edge angles, guide operative planning but are limited on radiographs by projection error and patient positioning 21,22,23,24 . CT is therefore widely used for quantitative morphometry and automated angle extraction has been developed for CT 25 . Automation of CT morphometry avoids the excess time-consumption and observer bias of manual measurements 26,27,28 , however, adjunct CT delivers ionizing radiation to a population of primarily young athletes and women of reproductive age 1,6 . Zero echo time (ZTE) MRI provides cortical bone contrast analogous to CT without ionizing radiation 20,29 , and manual ZTE morphometry agrees with CT 1 . ZTE MRI could provide comprehensive single-exam morphometric and soft-tissue evaluation, however, fully automated angle extraction has not been established for ZTE MRI. Prior automated MRI approaches rely on statistical shape modeling 30,31,32 , which may degrade with atypical anatomy. 4

Therefore, the purposes of this study were to (1) develop a fully automated pipeline that segments the femur, pelvis, and osseous landmarks from pelvic ZTE MRI and computes alpha, femoral neck-shaft, Tönnis, coronal and sagittal center-edge, and acetabular version angles using geometric modeling; and (2) evaluate agreement of automated angles with expert manual measurements. It was hypothesized that automated angles would agree with independent manual measurements from two experienced musculoskeletal radiologists at a level comparable to interrater reliability.

2

Methods

2.1

Enrollment

This prospective and retrospective single-institution cross-sectional diagnostic accuracy study (Level of Evidence III) was approved by the local Institutional Review Board (IRB# 2015441 and 2025-1851) and conducted in compliance with Health Insurance Portability and Accountability Act (HIPAA). Prospective participants enrolled with written informed consent. Retrospective participants were identified from clinically acquired pelvic MRI under a waiver of informed consent and HIPAA authorization. Inclusion required pelvic ZTE MRI with adequate image quality for manual segmentation; no participants were excluded after enrollment. A subset overlap with a prior publication evaluating manual ZTE–CT agreement 1 , which did not include automated angle computation. Sample size necessary for a held-out test set was estimated prospectively using Fisher’s z-transformation 33 for a one-sided test (H0 : ICC = 0; α = 0.05), with Breighner et al. 1 ZTE interrater ICC values as anticipated effect sizes. A minimum of 23 hips was required for 95% power for all eight angles.

5

2.2

Image Acquisition

Pelvic ZTE MRI was acquired on a clinical 3.0 T scanner (SIGNATM Premier, GE Healthcare, Waukesha, WI) with either a combination of a 30-channel adaptive image receive (AIR) anterior array with a 60-channel embedded AIR posterior array coil or 32-channel body array coil for larger patients, or a 32-channel cardiac coil for smaller patients (GE Healthcare, Waukesha, WI). Patients were positioned feet-first supine with the feet oriented neutrally and secured. ZTE was acquired following routine clinical pelvis MRI. Participants enrolled between 2017 and 2019 received prototype ZTE sequences post-processed as described by Breighner et al. 1 ; subsequent participants received the FDA-approved oZTEo sequence (GE Healthcare, Waukesha, WI) (Figure 1a) with proprietary shading correction. Images were acquired in the axial or coronal plane. The acquisition parameters for the prototype ZTE sequences were as follows: echo time, 0 ms; repetition time, 425–528 ms; flip angle, 1°; receiver bandwidth, ±62.5 kHz; number of excitations, 4; field of view, 36–44 cm; slice thickness, 1.1–1.5 mm; number of slices, 100–200; acquisition matrix, 320×320; and scan time, ∼5 minutes. The acquisition parameters for the oZTEo sequence were as follows: echo time, 0.016 ms; repetition time, 504.73 ms; flip angle, 1°; receiver bandwidth, ±83.33 kHz; field of view, 32–38 cm; slice thickness, 1.0–1.3 mm; number of slices, 180–320; acquisition matrix, 320 × 320; and scan time, ∼5–7 minutes.

2.3

Manual Angular Measurements

Standard measurements of hip morphology were derived for held-out test hips by boardcertified musculoskeletal radiologists with over 20 and 10 years of experience (YOE), radiologist 1 (E.A.B.) and radiologist 2 (M.S.), respectively. Angles measured were alpha, femoral neck-shaft, coronal and mid-acetabular sagittal center-edge, Tönnis, and acetabular version (at 1, 2, and 3 o’clock positions) angles. These measurements constitute the local clinical standard for assessment of FAI morphology with CT 1 . Images were provided within the radiologists’ native picture archiving and communication system (PACS) viewer, Sectra 6

Figure 1: Representative zero echo time (ZTE) hip magnetic resonance images (MRI) oZTEo acquisitions and manual segmentation of pelvis and femur. (a) Clinical oZTEo sequence acquired in axial (top) or coronal (bottom) orientation without manual post-processing. Displayed images received windowing to mimic CT contrast: level 1026 / window 1080. (b) Manual segmentation of femur (red) and pelvis (green). (c) Three fiducial landmarks were placed: lateral acetabulum (blue), medial weight-bearing acetabulum (yellow), and distal greater trochanter (cyan). The model was trained to predict femoral and pelvic masks plus these landmarks.

7

IDS7 (Sectra Medical Systems, Linköping, Sweden). All angles were measured to the nearest 0.1°. Manual measurements served as the reference standard for clinical validation of the automated pipeline.

2.4

Automated Segmentation Model Development

2.4.1

Manual Segmentation of Femur and Pelvis

The entire bone (trabecular and cortical) of the femur and pelvis was manually segmented in ITK-SNAP (v3.8.1) 34 from all hips by a biomedical engineer with greater than 3 YOE (J.C.) and reviewed as needed by a biomedical engineer with greater than 18 YOE (M.F.K.) and a board-certified radiologist with greater than 30 YOE (H.G.P.) (Figure 1b). High bone-tosoft-tissue contrast on ZTE images facilitated delineation. Bilateral exams were split at the midline so each hip could be assigned independently to training, validation, or test sets and to support acetabular version measurement.

2.4.2

Manual Identification of Fiducial Landmarks

Three fiducial landmarks were placed as separate labels within the bone masks: lateral acetabulum, medial weight-bearing acetabulum, and distal greater trochanter (J.C.; reviewed by E.A.B.) (Figure 1c). Landmarks were dilated to uniform 5-mm spheres.

2.4.3

Auto-segmentation Model Training

Data Preparation: Segmentation was performed using no-new convolutional network for biomedical image segmentation (nnU-Net) 35 . Digital imaging and communications in medicine (DICOM) volumes were converted to neuroimaging informatics technology initiative (NIfTI) format, standardized to left-posterior-inferior orientation, and cropped to the side of interest. Six labels were defined: background (0), femur (1), pelvis (2), lateral acetabulum (3), medial weight-bearing acetabulum (4), and distal greater trochanter (5).

8

Training: The default nnU-Net 3D full-resolution configuration was used. Training employed a weighted composite of cross-entropy and Dice loss with class weights of 3 for femur and 15 for fiducial labels relative to pelvis. Five-fold cross validation with an 80/20 train/validation split (per fold) was employed. Training continued until femur and pelvis mean validation Dice exceeded 0.90 and fiducial Dice exceeded 0.75 (or overall Dice exceeded 0.80), or 1000 epochs. Training and inference were performed on a single NVIDIA Tesla T4 GPU (16 GB VRAM) on an AWS Elastic-Compute Cloud g4dn.4xlarge instance running Ubuntu 22.04 with 4 CPUs across 2 cores, 125 GB of RAM, and CUDA-enabled (v12.8). Inference: Final segmentations were generated by using an ensemble of all five-fold-specific models using sliding-window prediction with a tile step size of 0.25 (nnU-Net default, 0.5), Gaussian fusion, and mirror-based test-time augmentation. If fiducial labels were absent from the ensemble mask, fold-specific predictions were evaluated for anatomically plausible placement (e.g., distal greater trochanter in contact with femur) before imputation into the final mask. Predictions received automated post-processing island removal for femur and pelvic labels, with only the largest connected regions retained.

2.5

Automated Angle Computation

All angles were computed in world coordinates (millimeters) from predicted segmentations. All described calculations assume standard Left-Posterior-Inferior-oriented NIfTI data converted to NumPy arrays via sitk.GetArrayFromImage() 36 . The axial (Superior–Inferior) axis, coronal (Anterior–Posterior) axis, and sagittal (Right–Left) axis are identified at runtime from the image direction cosine matrix by finding the column most aligned with the corresponding world directions.

2.5.1

Landmark and Center-of-Mass Localization

The center of mass of each predicted label was calculated in voxel space using scipy.ndimage.center of mass 37 and converted to world coordinates using the image affine 9

defined by origin, spacing, and direction cosines.

2.5.2

Femoral Reference Points

Two femoral reference points were derived from the femur mask (label 1) and distal greater trochanter landmark (label 5). The femur canal center is identified on the greater trochanter axial slice. The distal shaft center is identified on the most distal axial slice in the distal 5% of the volume whose femur voxel count was ≥95% of the mean count across a 10-slice sliding window. A conditional fallback was in place requiring the most distal slice contains >10 femur voxels. Slice-wise 2D centers of mass were converted to world coordinates.

2.5.3

Femoral Head Center

The femoral head center was estimated by robust Random Sample Consensus (RANSAC) sphere fitting to the proximal femoral surface.

The femur was divided at the greater

trochanter; the proximal half closer in axial position to the medial acetabulum (label 4) was retained. Surface points were obtained by binary erosion. Candidate spheres were fit from random 4-point subsets over 500 iterations with 3-mm inlier tolerance. The best model was refined by nonlinear least squares on inliers.

2.5.4

Automated Angle Definitions

Angles were defined by the angle between vector pairs in world space using the dot product. Femoral neck-shaft angle: Angle between the neck axis (greater trochanter to femoral head center) and shaft axis (greater trochanter to distal femur center). Tönnis angle: Angle between the inter-acetabular line and a horizontal reference constructed at the lateral acetabulum (label 3) coronal/sagittal coordinates and medial acetabulum (label 4) axial coordinate; signed inferiorly when the lateral acetabulum lay below the medial acetabulum.

10

Coronal center-edge angle: Angle between the femoral head center-to-lateral acetabulum vector and a vertical reference at the lateral acetabulum axial level and femoral head coronal/sagittal coordinates. Mid-acetabular sagittal center-edge angle: In the sagittal plane through the femoral head center, an anterior-inferior pelvic guide point was identified on the femoral-head sagittal slice. From the head center, 181 rays spanning 0° to 80° were cast in the sagittal plane; the sourcil was the first pelvic contact on the ray best aligned with the guide direction (largest vertical angle used to break ties). The angle was measured between the superior-inferior axis and the head-to-sourcil vector. Acetabular version (1, 2, 3): Axial slice levels were placed at clock-face positions 30°, 60°, and 90° from the femoral head equator toward the superior pole. At each level, posteriorlateral and anterior-lateral acetabular cusps were identified after splitting the pelvis mask at the femoral head coronal index. Version was measured relative to a reference line orthogonal to the line connecting ipsilateral and contralateral posterior pelvic points; retroversion was assigned a negative value. Alpha angle: Radial planes containing the femoral neck axis were rotated in 2° increments about the neck. On each plane, anterior head-neck junction points were identified along the femoral surface mesh; the plane with the lowest junction-band curvature variation was selected. The control point was the farthest anterior junction-band point from the head center, with optional distal extension when a head-cap gate indicated spherical-cap clustering. Alpha was the angle between the neck axis and the head-center-to-control-point vector.

2.6

Statistical Analysis

2.6.1

Model Performance

Held-out test segmentations were evaluated with Dice similarity coefficient, Jaccard index, 95th-percentile Hausdorff distance, and relative volume difference versus ground truth. Re11

sults are reported as mean ± standard deviation [minimum, maximum] per structure.

2.6.2

Fiducial Localization

Landmark center-point error was defined as the 3D Euclidean distance (mm) between predicted and ground-truth centers for each fiducial label; femoral head center error used the RANSAC-derived head center. Error is summarized by landmark.

2.6.3

Clinical Validation

Angles derived from manual measurements from two board-certified musculoskeletal radiologists (Rater 1 = E.A.B. and Rater 2 = M.S.) were compared with automated ensemble angles on test hips. Reliability was assessed with two-way random-effects, single-measurement, absolute-agreement ICC for inter-rater (Rater 1 versus Rater 2), intra-rater (Rater 1 repeat reads), and model versus rater-mean agreement. ICC was interpreted as: <0.40 poor, 0.40– 0.59 fair, 0.60–0.74 good, 0.75–1.00 excellent 38 , with statistical significance set at α = 0.05. Bland–Altman analysis 39 provided comparison of bias (mean difference) and 95% limits of agreement, computed as 1.96 times the standard deviation of differences (SDdiff ) for interrater (Rater 1 versus Rater 2) and model versus rater-mean. Difference was defined as Rater 1 − Rater 2 and Model − mean(Rater 1, Rater 2), respectively. Statistical analyses were performed in Python using SciPy (v1.13.1), NumPy (v1.26.4), pandas (v2.3.3), and Pingouin (v0.6.0).

3

Results

3.1

Participant Demographics

In total, 73 participants (51 women, mean ± SD [range] age: 36.8 ± 17.7 [14.5–82.9] years, BMI: 23.2 ± 4.0 [17.5–40.3] kg/m2 ; men: age 38.4 ± 20.6 [15.9–76.8] years, BMI: 28.2 ± 5.9 [19.6–42.5] kg/m2 ) and 135 hips were included (Table 1). One hundred hips were allocated to 12

model training and cross-validation. Manual angle validation was performed on 35 held-out test hips by two musculoskeletal radiologists; one radiologist repeated measurements on 10 test hips for intra-rater assessment. Table 1: Participant Demographics Metric

N (%)

Female Male

73 36.8 ± 18.5 [14.5–82.9] 51 (70.0) 36.2 ± 17.7 [14.5–82.9] 22 (30.0) 38.4 ± 20.6 [15.9–76.8]

Female Male

73 51 (70.0) 22 (30.0)

Hispanic or Latino Not Hispanic or Latino Unknown

73 4 (5.5) 66 (90.0) 3 (4.5)

Age

BMI

Ethnicity

Race

Mean ± SD [Range]

Sex

24.7 ± 5.2 [17.5–42.5] 23.2 ± 4.0 [17.5–40.3] 28.2 ± 5.9 [19.6–42.5]

73 American Indian or Alaska Native 0 (0.0) Asian 3 (2.7) Black or African American 1 (1.4) Native Hawaiian or Other Pacific Islander 0 (0.0) White 61 (83.6) Other/Unknown 9 (12.3)

Enrollment Prospective Retrospective

73 37 (50.7) 36 (49.3)

Unilateral Bilateral

73 9 (12.3) 64 (87.7)

Prototype oZTEo

73 23 (31.5) 50 (68.5)

Bilateral Exam

ZTE Acquisition

Note.—N = number of participants, BMI = body mass index. Age and BMI are displayed as mean ± standard deviation [range]; N is displayed as count (%).

3.2

Segmentation Performance

On the held-out test set, femur and pelvis segmentation achieved high overlap (Dice 0.98 ± 0.01 and 0.97 ± 0.01, respectively) (Table 2). Fiducial labels had lower but acceptable over13

lap (lateral acetabulum 0.83 ± 0.09; medial acetabulum 0.65 ± 0.18; distal greater trochanter 0.66 ± 0.22). Jaccard index was consistent with Dice, with excellent overlap for bone (femur 0.96 ± 0.01; pelvis 0.94 ± 0.02) and moderate overlap for fiducials (0.71 ± 0.12 lateral acetabulum, 0.51 ± 0.19 medial acetabulum, 0.52 ± 0.20 distal greater trochanter). Relative volume difference was greatest for femur and pelvis (3.66 ± 2.13 cm3 and 4.50 ± 3.11 cm3 , respectively), whereas fiducial volume differences remained an order of magnitude smaller (approximately 0.12–0.15 cm3 ), apart from outlier cases with incomplete fiducial predictions. Mean femur and fiducial Hausdorff distances remained 1–4 mm, reflecting small surface offsets on compact landmarks and the more linear femur geometry. The mean Hausdorff of the pelvis was 12.2 mm, with greater offset for larger geometries. Table 2: Auto-segmentation model performance Label 1 2 3 4 5

Structure

n

Dice

Femur 35 0.98 ± 0.01 Pelvis 35 0.97 ± 0.01 Lateral Acetabulum 35 0.83 ± 0.09 Medial Acetabulum 35 0.65 ± 0.18 Distal Greater Trochanter 35 0.66 ± 0.22

Jaccard Index

Hausdorff Distance (mm)

Volume Difference (cm3 )

0.96 ± 0.01 [0.93, 0.98] 0.94 ± 0.02 [0.87, 0.97] 0.71 ± 0.12 [0.40, 0.92] 0.51 ± 0.19 [0.10, 0.84] 0.52 ± 0.20 [0.02, 0.84]

3.5 ± 1.4 [1.4, 6.7] 12.2 ± 10.4 [3.0, 56.6] 2.0 ± 0.8 [1.0, 5.0] 3.5 ± 1.6 [1.0, 7.8] 3.6 ± 2.4 [1.4, 11.2]

3.66 ± 2.13 [0.83, 9.13] 4.50 ± 3.11 [0.76, 13.25] 0.12 ± 0.99 [0.01, 0.39] 0.15 ± 0.13 [0.01, 0.63] 0.15 ± 0.15 [0.00, 0.73]

Note.—n = number of hips, mm = millimeters. All metrics are relative comparisons of ground truth to predicted segmentations before any prediction clean-up. Cross-validation mean pseudo Dice and test set Dice score reported as mean ± standard deviation. Jaccard index, 95th percentile Hausdorff distance, and relative volume difference reported as mean ± standard deviation [range]. A maximum of 35 hips for each label were possible within the validation and test set, respectively.

3.3

Fiducial Localization

Center-point error was smallest for the femoral head (0.38 ± 0.18 mm; median 0.38 mm) and lateral acetabulum (1.03 ± 0.78 mm; median 0.82 mm). The medial acetabulum and distal greater trochanter had greater error (2.46 ± 1.29 mm and 2.54 ± 1.77 mm; medians 2.48 and 1.97 mm, respectively), with wider ranges when trochanter morphology was indistinct.

14

3.4

Clinical Validation

3.4.1

Interrater Agreement

Acetabular versions 1 (ICC: 0.82), 2 (ICC: 0.86), and 3 (ICC: 0.92), coronal center-edge (ICC: 0.93), and Tönnis angle (ICC: 0.93) showed excellent interrater reliability (p < 0.001) (Table 3 and Figure 2, Column 1). Mid-acetabular sagittal center-edge had fair ICC (0.45, p < 0.001). Alpha (ICC: 0.20, p = 0.129) and femoral neck-shaft angle (ICC: 0.20, p = 0.051) had poor reliability. Acetabular versions 1, 2, and 3, coronal center-edge, and Tönnis angle had small biases of +2.67°, +1.11°, +0.43°, +1.29°, and −0.12°, respectively. Acetabular version 3, coronal center-edge, Tönnis angle had smaller 1.96 × SDdiff of 2.48°–3.03° than acetabular versions 1 and 2 with 1.96 × SDdiff of 4.53° and 3.59°, respectively. Mid-acetabular sagittal centeredge had large systematic disagreement (−7.91 ± 7.71°). The widest limit of agreement was observed for alpha (11.54°), though only a small bias −1.62° was observed. Femoral neck-shaft angle had moderate disagreement (+3.18 ± 4.33°). Table 3: Interrater agreement assessed via ICC and Bland–Altman

Angle

n

Intraclass Correlation Coefficient

Bland–Altman (°)

ICC [95% CI]

Bias ± 1.96×SDdiff

p

Reliability

Alpha 35 0.20 [−0.14, 0.49] 0.129 Poor Acetabular version 1 35 0.82 [0.58, 0.92] <0.001 Excellent Acetabular version 2 35 0.86 [0.74, 0.93] <0.001 Excellent Acetabular version 3 35 0.92 [0.85, 0.96] <0.001 Excellent Coronal center-edge 35 0.93 [0.85, 0.97] <0.001 Excellent Femoral neck-shaft 35 0.20 [−0.07, 0.47] 0.051 Poor Mid-acetabular sagittal CE 35 0.45 [−0.04, 0.73] <0.001 Fair Tönnis 35 0.93 [0.86, 0.96] <0.001 Excellent

−1.62 ± 11.54 +2.67 ± 4.53 +1.11 ± 3.59 +0.43 ± 2.48 +1.29 ± 3.03 +3.18 ± 4.33 −7.91 ± 7.71 −0.12 ± 2.48

Note.—ICC = intraclass correlation coefficient, SD = standard deviation, CE = center-edge. ICC reported as point estimate [95% confidence interval]. Bland–Altman: bias ± 1.96×SDdiff in degrees.

15

Figure 2: Bland–Altman agreement for automated and manual hip angle measurements on zero echo time (ZTE) magnetic resonance images (MRI). Rows correspond to individual angular measures; two columns compare Rater 1 (E.A.B.) with Rater 2 (M.S.) and the model with the mean of Rater 1 and Rater 2. Points show per-hip differences from the mean paired value; horizontal lines denote mean bias (teal, solid) and ±1.96 × SDdiff (light teal, dashed), with the gray band spanning one 1.96 × SDdiff . The zero line is indicated as a thin black dashed line. Tighter point clouds indicate closer agreement; alpha and sagittal center-edge at mid femoral head height show the greatest dispersion. 16

3.4.2

Intra-rater Agreement

Acetabular version 3 (ICC: 0.96), coronal center-edge (ICC: 0.96), and Tönnis angle (ICC: 0.95) remained excellent on 10 repeated reads by Rater 1 (all p < 0.001), with small biases (+1.31°, +1.72°, and −0.54°) and 1.96 × SDdiff of 1.84°–2.58° (Table 4). Acetabular version 1 showed good reliability (ICC: 0.83, p = 0.001; bias ±1.96×SDdiff : +1.40±4.05°). Acetabular version 2 (ICC: 0.59, p = 0.020; +3.13±6.47°) and mid-acetabular sagittal center-edge (ICC: 0.56, p = 0.013; −7.51 ± 9.61°) were fair, though mid-acetabular sagittal center-edge had a large bias and limits of agreement. Alpha (ICC: −0.18, p = 0.686) and femoral neck-shaft angle (ICC: 0.26, p = 0.222) showed poor reliability, with alpha displaying the widest limits of agreement (+3.50 ± 16.70°) and femoral neck-shaft angle moderate spread (+1.53 ± 5.19°). Table 4: Intra-rater agreement assessed via ICC and Bland–Altman

Angle

n

Alpha 10 Acetabular version 1 10 Acetabular version 2 10 Acetabular version 3 10 Coronal center-edge 10 Femoral neck-shaft 10 Mid-acetabular sagittal CE 10 Tönnis 10

Intraclass Correlation Coefficient

Bland–Altman (°)

ICC [95% CI]

Bias ± 1.96×SDdiff

−0.18 [−0.78, 0.50] 0.83 [0.49, 0.96] 0.59 [0.05, 0.88] 0.96 [0.81, 0.99] 0.96 [0.78, 0.99] 0.26 [−0.40, 0.75] 0.56 [−0.02, 0.87] 0.95 [0.83, 0.99]

p

Reliability

0.686 Poor <0.001 Excellent 0.020 Fair <0.001 Excellent <0.001 Excellent 0.222 Poor 0.013 Fair <0.001 Excellent

+3.50 ± 16.70 +1.40 ± 4.05 +3.13 ± 6.47 +1.31 ± 1.84 +1.72 ± 2.39 +1.53 ± 5.19 −7.51 ± 9.61 −0.54 ± 2.58

Note.—ICC = intraclass correlation coefficient, SD = standard deviation, CE = center-edge. ICC reported as point estimate [95% confidence interval]. Bland–Altman: bias ± 1.96×SDdiff in degrees.

3.4.3

Model vs. Rater-Mean Agreement

Automatically computed acetabular versions 1–3, coronal center-edge, and Tönnis showed excellent agreement with rater mean (ICC: 0.92–0.96, all p < 0.001), with small biases (−1.64° to +1.11°) and narrower limits of agreement than interrater (1.85°–2.77°) (Table 5 and Figure 2, Column 2). Automated mid-acetabular sagittal center-edge showed good agreement (ICC: 0.74, p < 0.001; bias ±1.96 × SDdiff : −1.75 ± 5.45°). ICC for automated Alpha (0.45, p = 0.003) and femoral neck-shaft (0.55, p < 0.001) were improved 17

from interrater ICC. Femoral neck-shaft had a smaller bias and tighter limits of agreement (+1.88 ± 3.75°) while automated alpha had bias ±1.96 × SDdiff consistent with the interrater analysis (−1.91 ± 8.32°). Table 5: Model vs. mean(Rater 1, Rater 2) agreement assessed via ICC and Bland–Altman Intraclass Correlation Coefficient

Bland–Altman (°)

n

ICC [95% CI]

Bias ± 1.96×SDdiff

Alpha 35 Acetabular version 1 35 Acetabular version 2 35 Acetabular version 3 35 Coronal center-edge 35 Femoral neck-shaft 35 Mid-acetabular sagittal CE 35 Tönnis 35

0.45 [0.15, 0.68] 0.93 [0.80, 0.97] 0.96 [0.91, 0.98] 0.95 [0.91, 0.98] 0.95 [0.87, 0.98] 0.55 [0.24, 0.75] 0.74 [0.54, 0.86] 0.92 [0.82, 0.96]

Angle

p

Reliability

0.003 Fair <0.001 Excellent <0.001 Excellent <0.001 Excellent <0.001 Excellent <0.001 Fair <0.001 Good <0.001 Excellent

−1.91 ± 8.32 −1.64 ± 2.77 +0.77 ± 1.85 +0.48 ± 1.90 −1.35 ± 2.24 +1.88 ± 3.75 −1.75 ± 5.45 +1.11 ± 2.17

Note.—ICC = intraclass correlation coefficient, SD = standard deviation, CE = center-edge. ICC reported as point estimate [95% confidence interval]. Bland–Altman: bias ± 1.96×SDdiff in degrees.

4

Discussion

This study demonstrates that a fully automated ZTE MRI pipeline combining nnU-Net segmentation, fiducial landmark detection, and geometric modeling can derive hip morphometric angles highly correlated with dual-expert manual measurements. This is the first study to compute alpha, femoral neck-shaft, Tönnis, coronal and sagittal center-edge, and multiple acetabular version angles directly from ZTE MRI using deep learning coupled to explicit landmark-based geometry rather than shape inference alone. The pipeline improves upon previous MRI-native methods that rely on statistical shape models 30,31,32 that may degrade in the setting of marked deformity. The fiducial-first approach embeds radiologistintended anatomic priors while retaining full automation, enabling automated review to function closer to an additional rater rather than a population-based estimation. Accurate bone segmentation and submillimeter to low single-digit millimeter fiducial center-point error support bone and fiducial landmark-based angle computation. Inter-rater 18

reliability using the automated landmarks was excellent for acetabular version, Tönnis, and coronal coverage measures. Model versus mean grades had improved agreement over expert manual interrater data for mid-acetabular sagittal center-edge, indicating this standardized, automated method may produce more reliable results. Poor alpha and femoral neck-shaft measurements are consistent with previous reporting 27,28,32,40,41 and model versus mean rater consensus was within the inter- and intra-rater agreement range. Evaluating the magnitude of difference for femoral neck-shaft, poor agreement did not indicate clinically significant differences. The femoral neck-shaft angle has a normal range of 120° to 135°, thus a 3° to 5° variation within this range may be inconsequential 1,26 . The alpha angle was the most challenging measure because manual identification of the radial cam-defining plane is inherently variable 27,28,32 . The automated method selects the flattest junction-band radial section under explicit rules, yielding a reproducible estimate of maximal cam prominence even when numeric agreement with a specific reader is limited. When combined with standard of care pelvis MRI, automated ZTE morphometry could streamline preoperative evaluation of hip-preservation candidates by consolidating osseous and soft-tissue assessment in one study, obviating the need for a CT scan and time-consuming manual hip angle measurements. Post-operative imaging could also be performed with automated measurements to improve reliability of measurements between examinations, which previously may have been performed by raters with varying experience and techniques. As a radiation-sparing examination, credible FAI morphometry within MRI enables a greater ability to safely and accurately track FAI progression in relatively young athletes across seasons. Areas for further study include determining if the use of automated angles improves radiologist and referring physician confidence in measurements, as well as cost and time savings over the current multi-modality imaging workflow. This study had several limitations. Data were acquired at a single institution using a single ZTE acquisition approach while other sites and MRI vendors may utilize other ZTE acquisition and reconstruction techniques 42 . We anticipate that our methods are fully trans-

19

latable if images generated with an alternative sequence undergo semantic segmentation to define the femur and acetabulum. Second, the cohort was not demographically or athletically characterized in detail, and deformity subtype was not systematically classified. Third, most fiducial training labels were placed by a trained research engineer under radiologist oversight rather than annotated directly by a radiologist. Fourth, it may be anticipated that landmark performance will degrade with severe deformity, artifact, or atypical anatomy, particularly at the medial acetabulum. Further, alpha and sagittal center-edge depend on algorithmic choices that may not mirror manual convention. We anticipate that continued enrollment and additional model training will be able to account for natural variability in bone shape. Fifth, the intra-rater assessment was performed for only 10 hips, although our data is similar to previous experience 1 . Additional reads may change agreement statistics, although minor changes are expected given similar intrarater performance previously reported 1 . Finally, the angles were evaluated without family-wise error control. In conclusion, this study validated a fully automated ZTE MRI pipeline that segments bone and fiducial landmarks and computes eight hip angles commonly used to evaluate patients with FAI and hip dysplasia. Model versus rater-mean agreement was excellent for acetabular versions 1–3, coronal center-edge, and Tönnis (ICC: 0.92–0.96), with Bland– Altman limits comparable to or narrower than interrater limits for most version and coverage angles. Alpha and femoral neck-shaft angle require cautious interpretation given fair model versus mean-rater and poor interrater reliability. This approach supports efficient, radiationsparing, single-modality bone morphometric and soft tissue hip evaluation in hip-preservation patients.

Acknowledgements The authors would like to thank the staff of the Hospital for Special Surgery Department of Radiology and Imaging for their assistance with scanning and acquisition of image data.

20

The authors used Claude (Anthropic, San Francisco, CA), a large language model, to assist with editing and formatting of this manuscript; all scientific content, data analysis, interpretation of results, and conclusions are solely the work of the authors. The authors take full responsibility for the accuracy and integrity of the submitted work.

Funding Statement The authors received no financial support for the research, authorship, and/or publication of this article.

Ethical Statement This study was approved by the local Institutional Review Board (IRB# 2015-441 and 2025-1851) and conducted in compliance with the Health Insurance Portability and Accountability Act (HIPAA). Prospective participants enrolled with written informed consent. Retrospective participants were identified from clinically acquired pelvic MRI under a waiver of informed consent and HIPAA authorization.

Data Sharing Statement Data generated or analyzed during the study are available from the corresponding author by request.

Author Contributions Jack Consolini: Conceptualization, Data curation, Formal analysis, Investigator, Methodology, Project administration, Software, Supervision, Statistical Analysis, Validation, Visualization, Writing – original draft, Writing – review and editing.

21

Eric A. Bogner: Formal analysis, Investigation, Methodology, Validation, Writing – review and editing. Meghan Sahr: Formal analysis, Investigation, Methodology, Validation, Writing – review and editing. Matthew F. Koff: Conceptualization, Data Curation, Investigation, Project administration, Resources, Supervision, Writing – review and editing. Kevin M. Koch: Conceptualization, Funding acquisition, Project administration, Resources, Supervision, Visualization, Writing – review and editing. Hollis G. Potter: Conceptualization, Funding acquisition, Project administration, Resources, Supervision, Visualization, Writing – review and editing.

References [1] R. E. Breighner, E. A. Bogner, S. C. Lee, M. F. Koff, and H. G. Potter. Evaluation of Osseous Morphology of the Hip Using Zero Echo Time Magnetic Resonance Imaging. Am J Sports Med, 47(14):3460–3468, 2019. [2] D. R. Griffin, E. J. Dickenson, J. O’Donnell, et al. The Warwick Agreement on femoroacetabular impingement syndrome (FAI syndrome): an international consensus statement. Br J Sports Med, 50(19):1169–1176, 2016. [3] R. F. Hale, H. P. Melugin, J. Zhou, et al. Incidence of Femoroacetabular Impingement and Surgical Management Trends Over Time. Am J Sports Med, 49(1):35–41, 2021. [4] M. E. Dancy, V. Oladipo, P. Boadi, et al. Femoroacetabular Impingement: Critical Analysis Review of Current Nonoperative Treatments. JBJS Rev, 13(4):e24.00211, 2025. [5] C. Zhang, L. Li, B. B. Forster, et al. Femoroacetabular impingement and osteoarthritis of the hip. Can Fam Physician, 61(12):1055–1060, 2015.

22

[6] M. Tanzer and N. Noiseux. Osseous Abnormalities and Early Osteoarthritis: The Role of Hip Impingement. Clin Orthop Relat Res, (429):170, 2004. [7] P. Lorenzon, A. Scalvi, and E. Scalco. The Painful Hip in Young Adults Between Impingement and Mild Dysplasia: Clinical and Instrumental Diagnostical Criteria. Acta Bio Medica, 91(4-S):11–20, 2020. [8] R. T. Loder and E. N. Skopelja. The Epidemiology and Demographics of Hip Dysplasia. ISRN Orthop, 2011:238607, 2011. [9] M. Beck, M. Kalhor, M. Leunig, and R. Ganz. Hip morphology influences the pattern of damage to the acetabular cartilage: femoroacetabular impingement as a cause of early osteoarthritis of the hip. J Bone Joint Surg Br, 87(7):1012–1018, 2005. [10] A. E. Li, S. T. Jawetz, H. G. Greditzer, A. J. Burge, D. H. Nawabi, and H. G. Potter. MRI for the preoperative evaluation of femoroacetabular impingement. Insights Imaging, 7(2):187–198, 2016. [11] M. J. Philippon, C. P. Ho, K. K. Briggs, J. Stull, and R. F. LaPrade. Prevalence of increased alpha angles as a measure of cam-type femoroacetabular impingement in youth ice hockey players. Am J Sports Med, 41(6):1357–1362, 2013. [12] S. D. Ferrell, S. X. Ramirez, S. B. Rosenfeld, P. X. Bawa, A. C. Sher, and J. H. Kan. Does presence of hip dysplasia affect the prevalence and location of cam morphologies in pediatric hip pain patients? Clin Imaging, 131:110731, 2026. [13] C. Y. W. Heimer, C. H. Wu, C. Perka, et al. The Impact of Hip Dysplasia on CAM Impingement. J Pers Med, 12(7):1129, 2022. [14] K. A. Siebenrock, F. Ferner, P. C. Noble, R. F. Santore, S. Werlen, and T. C. Mamisch. The cam-type deformity of the proximal femur arises in childhood in response to vigorous sporting activity. Clin Orthop, 469(11):3229–3240, 2011. 23

[15] A. K. Yépez, M. Abreu, B. Germani, and C. R. Galia. Prevalence of femoroacetabular impingement morphology in asymptomatic youth soccer players. Rev Bras Ortop, 52(Suppl 1):14–20, 2017. [16] C. J. Guevara, R. Pietrobon, J. T. Carothers, S. A. Olson, and T. P. Vail. Comprehensive morphologic evaluation of the hip in patients with symptomatic labral tear. Clin Orthop, (453):277–285, 2006. [17] C. R. Henak, C. L. Abraham, A. E. Anderson, et al. Patient-specific analysis of cartilage and labrum mechanics in human hips with acetabular dysplasia. Osteoarthritis Cartilage, 22(2):210–217, 2014. [18] A. E. Li, S. T. Jawetz, H. G. Greditzer, A. J. Burge, D. H. Nawabi, and H. G. Potter. MRI Evaluation of Femoroacetabular Impingement After Hip Preservation Surgery. AJR Am J Roentgenol, 207(2):392–400, 2016. [19] S. L. Gold, A. J. Burge, and H. G. Potter. MRI of hip cartilage: joint morphology, structure, and composition. Clin Orthop, 470(12):3321–3331, 2012. [20] J. Goronzy, S. Blum, A. Hartmann, et al. Is MRI an adequate replacement for CT scans in the three-dimensional assessment of acetabular morphology?

Acta Radiol,

60(6):726–734, 2019. [21] K. M. Smith, B. J. Gerrie, P. C. McCulloch, D. M. Lintner, and J. D. Harris. Comparison of MRI, CT, Dunn 45° and Dunn 90° alpha angle measurements in femoroacetabular impingement. Hip Int, 28(4):450–455, 2018. [22] D. M. Falgout, P. J. Bevan, R. C. Grumet, and K. C. Parvaresh. Femoroacetabular Impingement Measurements Obtained From Two-Dimensional Radiographs Versus Three-Dimensional-Reconstructed Computed Tomography Images Result in Different Values. Arthrosc Sports Med Rehabil, 6(1):100833, 2023.

24

[23] T. Sakai, T. Nishii, K. Sugamoto, H. Yoshikawa, and N. Sugano. Is vertical-centeranterior angle equivalent to anterior coverage of the hip? Clin Orthop, 467(11):2865– 2871, 2009. [24] M. Tannast, K. A. Siebenrock, and S. E. Anderson. Femoroacetabular impingement: radiographic diagnosis—what the radiologist should know. AJR Am J Roentgenol, 188(6):1540–1552, 2007. [25] S. Tayyebinezhad, M. Fatehi, H. Arabalibeik, and H. Ghadiri. Intelligent extraction of CT image landmarks for improving cam-type femoroacetabular impingement assessment. Eur Radiol, 2025. [26] M. Sangeux, J. Pascoe, H. K. Graham, F. Ramanauskas, and T. Cain.

Three-

dimensional measurement of femoral neck anteversion and neck shaft angle. J Comput Assist Tomogr, 39(1):83–85, 2015. [27] M. R. Nouh, M. E. Schweitzer, L. Rybak, and J. Cohen. Femoroacetabular impingement: can the alpha angle be estimated? AJR Am J Roentgenol, 190(5):1260–1262, 2008. [28] H. P. Nötzli, T. F. Wyss, C. H. Stoecklin, M. R. Schmid, K. Treiber, and J. Hodler. The contour of the femoral head-neck junction as a predictor for the risk of anterior impingement. J Bone Joint Surg Br, 84(4):556–560, 2002. [29] E. G. Koretsky, D. Singh, J. Consolini, et al. Deep learning-enhanced zero echo time MRI for femoral cortical bone quantification: A technical validation against CT. Magn Reson Imaging, 133:110748, 2026. [30] Y. Xia, J. Fripp, S. S. Chandra, D. Walker, S. Crozier, and C. Engstrom. Automated 3D quantitative assessment and measurement of alpha angles from the femoral head-neck junction using MR imaging. Phys Med Biol, 60(19):7601–7616, 2015.

25

[31] J. M. Bugeja, Y. Xia, S. S. Chandra, et al. Automated volumetric and statistical shape assessment of cam-type morphology of the femoral head-neck region from clinical 3D magnetic resonance images. Quant Imaging Med Surg, 12(10):4924–4941, 2022. [32] N. P. Ewertowski, C. Schleich, D. B. Abrar, H. S. Hosalkar, and B. Bittersohl. Automated measurement of alpha angle on 3D-magnetic resonance imaging in femoroacetabular impingement hips: a pilot study. J Orthop Surg, 17(1):370, 2022. [33] J. Cohen. Statistical Power Analysis for the Behavioral Sciences. Routledge, 2nd edition, 2013. [34] P. A. Yushkevich, J. Piven, H. C. Hazlett, et al. User-guided 3D active contour segmentation of anatomical structures: Significantly improved efficiency and reliability. NeuroImage, 31(3):1116–1128, 2006. [35] F. Isensee, P. F. Jaeger, S. A. A. Kohl, J. Petersen, and K. H. Maier-Hein. nnu-Net: a self-configuring method for deep learning-based biomedical image segmentation. Nat Methods, 18(2):203–211, 2021. [36] B. C. Lowekamp, D. T. Chen, L. Ibáñez, and D. Blezek. The Design of SimpleITK. Front Neuroinformatics, 7:45, 2013. [37] P. Virtanen, R. Gommers, T. E. Oliphant, et al. SciPy 1.0: fundamental algorithms for scientific computing in Python. Nat Methods, 17(3):261–272, 2020. [38] D. V. Cicchetti. Guidelines, criteria, and rules of thumb for evaluating normed and standardized assessment instruments in psychology. Psychol Assess, 6(4):284–290, 1994. [39] J. M. Bland and D. G. Altman. Statistical methods for assessing agreement between two methods of clinical measurement. Lancet, 1(8476):307–310, 1986.

26

[40] E. Liodakis, I. Doxastaki, K. Chu, et al. Reliability of the assessment of lower limb torsion using computed tomography: analysis of five different techniques. Skeletal Radiol, 41(3):305–311, 2012. [41] K. L. Hermann and N. Egund. CT measurement of anteversion in the femoral neck. The influence of femur positioning. Acta Radiol, 38(4):527–532, 1997. [42] D. M. Grodzki, P. M. Jakob, and B. Heismann. Ultrashort echo time imaging using pointwise encoding time reduction with radial acquisition (PETRA). Magn Reson Med, 67(2):510–518, 2012.

27

Related documents

• Fast surface reconstruction of human brain MRI: benchmarking deep-learning based morphometry tools.#340425
• Automated Lesion Segmentation of Stroke MRI Using nnU-Net: A Comprehensive External Validation Across Acute and Chronic Lesions#1118461
• PRIME: Phase reversed interleaved multi-Echo acquisition enables highly accelerated distortion-corrected diffusion MRI.#1147273
• Pediatric pancreas segmentation from MRI scans with deep learning.#251309
• Comparative Study of Anatomical and Learned Features in AI Models for Structural Brain MRI#974894