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

Background and Rationale: The paper states that Computed tomography (CT) remains the reference for three-dimensional osseous morphometry for patients with femoroacetabular impingement (FAI) but entails ionizing radiation and manual measurement. The authors note that 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. The paper highlights that FAI and dysplasia of the hip disrupt normal joint biomechanics and are established contributors to early osteoarthritis in young and active patients, and that Pre-operative assessment relies on radiography for screening and CT for detailed three-dimensional osseous morphometry, while MRI remains essential for soft-tissue evaluation of the labrum and articular cartilage. Furthermore, adjunct CT delivers ionizing radiation to a population of primarily young athletes and women of reproductive age. Prior automated MRI approaches rely on statistical shape modeling, which may degrade with atypical anatomy.

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. The hypothesis was that automated angles would agree with independent manual measurements from two experienced musculoskeletal radiologists at a level comparable to interrater reliability.

Study Design: Cross-sectional diagnostic accuracy study, Level of Evidence 3.

Methods:

• Enrollment: 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. The study included 100 hips for training and 35 held-out test hips for validation. A subset overlap with a prior publication evaluating manual ZTE–CT agreement, which did not include automated angle computation. Sample size was estimated prospectively, requiring a minimum of 23 hips... for 95% power for all eight angles.

• Image Acquisition: MRI was performed on a clinical 3.0 T scanner (SIGNATM Premier, GE Healthcare) using prototype ZTE sequences (2017–2019) or the FDA-approved oZTEo sequence (GE Healthcare) with acquisition parameters including echo time of 0–0.016 ms, flip angle of 1°, slice thickness of 1.0–1.5 mm, and scan time of approximately 5–7 minutes.

• Manual Measurements: "Standard measurements of hip morphology were derived for held-out test hips by board-certified 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."

• Segmentation Model: An nnU-Net model was trained on 100 manually curated hips to automatically segment the femur, pelvis, and three osseous landmarks. Six labels were defined: background (0), femur (1), pelvis (2), lateral acetabulum (3), medial weight-bearing acetabulum (4), and distal greater trochanter (5). Training used the default nnU-Net 3D full-resolution configuration with a weighted composite of cross-entropy and Dice loss with class weights of 3 for femur and 15 for fiducial labels relative to pelvis and five-fold cross-validation. Inference used an ensemble of all five-fold-specific models using sliding-window prediction... Gaussian fusion, and mirror-based test-time augmentation.

• Geometric Analysis: Custom-developed automated geometric analysis algorithms computed alpha, femoral neck-shaft, Tönnis, coronal and sagittal center-edge, and acetabular version angles from inferred segmentations. Notable algorithmic features included RANSAC sphere fitting to the proximal femoral surface for femoral head center estimation and landmark/center-of-mass localization.

• Statistics: 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. ICC was interpreted as "<0.40 poor, 0.40–0.59 fair, 0.60–0.74 good, 0.75–1.00 excellent."

Results:

• 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). Dice coefficient exceeded 0.96 for bone segmentation (as stated in the abstract). Fiducial labels had lower but acceptable overlap (lateral acetabulum 0.83 ± 0.09; medial acetabulum 0.65 ± 0.18; distal greater trochanter 0.66 ± 0.22). Jaccard indices were 0.96 ± 0.01 for femur and 0.94 ± 0.02 for pelvis, with fiducials ranging from 0.51–0.71. Hausdorff distances were 3.5 ± 1.4 mm for femur and 12.2 ± 10.4 mm for pelvis.

• 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. The abstract summarizes: "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)."

• 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)." "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. Bland–Altman analysis showed small biases of +2.67°, +1.11°, +0.43°, +1.29°, and −0.12° for acetabular versions 1–3, coronal center-edge, and Tönnis, respectively, with the widest limit of agreement... observed for alpha (11.54°), though only a small bias −1.62° was observed."

• Intra-rater Agreement: On 10 repeated reads by Rater 1, Acetabular version 3 (ICC: 0.96), coronal center-edge (ICC: 0.96), and Tönnis angle (ICC: 0.95) remained excellent, whereas Alpha (ICC: −0.18) and femoral neck-shaft angle (ICC: 0.26) showed poor reliability, with alpha displaying the widest limits of agreement (+3.50 ± 16.70°).

• 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°)." "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 from interrater ICC. The abstract summarizes: 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."

Conclusion: Fully automated quantitative morphometric assessment from ZTE MRI is feasible and performs comparably to an expert reader for most coverage and version angles. The paper states that "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, radiation-sparing, single-modality bone morphometric and soft tissue hip evaluation in hip-preservation patients."

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. The paper also notes that 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."

Key Novelty: "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 fiducial-first approach embeds radiologist-intended anatomic priors while retaining full automation, enabling automated review to function closer to an additional rater rather than a population-based estimation."

Limitations: The authors acknowledge that 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; the cohort was not demographically or athletically characterized in detail, and deformity subtype was not systematically classified; most fiducial training labels were placed by a trained research engineer under radiologist oversight rather than annotated directly by a radiologist; landmark performance may degrade with severe deformity, artifact, or atypical anatomy, particularly at the medial acetabulum; the intra-rater assessment was performed for only 10 hips; and the angles were evaluated without family-wise error control.

• Improvement: Train segmentation networks with class weights scaled by clinical importance and inverse frequency (3× for femur, 15× for fiducials relative to pelvis), rather than uniform weighting.

• What the improved system can do: Segment tiny, clinically critical structures (lateral acetabulum, medial weight-bearing acetabulum, greater trochanter) with Dice 0.65–0.83 and sub-millimeter to 2.5 mm landmark localization, despite occupying a minuscule fraction of the volume — structures standard U-Nets typically miss or merge into background.

• Improvement: Replace end-to-end angle regression or statistical shape modeling with a multi-stage system: (a) nnU-Net segmentation, (b) RANSAC sphere fitting to the proximal femoral surface for head-center estimation, (c) center-of-mass and landmark localization, (d) deterministic geometric angle computation from these primitives.

• What the improved system can do: Achieve excellent agreement with expert raters (ICC 0.92–0.96 for acetabular version, coronal center-edge, and Tönnis angles) while remaining robust to atypical anatomy that degrades statistical shape models — because the geometry is derived from the actual patient scan rather than a population prior.

• Improvement: Embed radiologist-intended anatomic landmarks (lateral acetabulum, medial weight-bearing acetabulum, distal greater trochanter) as explicit output labels, so the model learns anatomically meaningful intermediate representations instead of latent features.

• What the improved system can do: Function like an additional human rater rather than a population-based estimator — its measurement behavior (biases, limits of agreement) is interpretable and comparable to inter-rater statistics, enabling clinicians to trust it as a peer reviewer rather than a black box.

• Improvement: Train/validate against the mean of multiple independent expert raters rather than a single rater's annotations, and evaluate with ICC plus Bland–Altman against that consensus.

• What the improved system can do: Outperform any single human rater in consistency — the model achieved narrower Bland–Altman limits of agreement (1.85°–2.77°) than inter-rater agreement, effectively averaging out human variability and producing more reproducible measurements than a typical clinical read.

• Improvement: Incorporate the empirical finding that some angles have inherently poor inter-rater reliability (alpha ICC 0.20, femoral neck-shaft ICC 0.20, sagittal center-edge ICC 0.45) into the system's reporting logic.

• What the improved system can do: Automatically classify each output angle into tiers — clinically reliable, fully automated (version 1–3, coronal center-edge, Tönnis), good, use with awareness (mid-acetabular sagittal center-edge), and flag for human review / cautious interpretation (alpha, femoral neck-shaft) — preventing false confidence in measurements where even expert readers fundamentally disagree.

• Improvement: Deploy a five-fold cross-validation ensemble with sliding-window prediction, Gaussian fusion, and mirror-based test-time augmentation.

• What the improved system can do: Achieve Dice >0.96 for whole-bone segmentation (femur 0.98, pelvis 0.97) on held-out patients, with the ensemble robustness needed for clinical deployment on unseen anatomy.

• Improvement: Integrate automated ZTE-MRI bone morphometry with standard MRI soft-tissue evaluation (labrum, cartilage) in one pipeline, eliminating the need for adjunct CT.

• What the improved system can do: Deliver a complete hip-preservation workup — eight osseous angles plus soft-tissue pathology — from a single 5–7 minute radiation-free MRI scan, reducing ionizing radiation exposure in young athletes and women of reproductive age, and eliminating time-consuming manual angle measurement.

• Improvement: Propagate segmentation uncertainty (e.g., landmark localization error of 0.38 mm femoral head, 1.03 mm lateral acetabulum, <2.5 mm medial acetabulum/greater trochanter) through the geometric angle computations to output per-angle confidence intervals rather than point estimates.

• What the improved system can do: Report that version and coverage angles carry sub-degree uncertainty (narrow limits of agreement), while flagging that alpha and femoral neck-shaft computations inherit higher uncertainty from both segmentation and inherent human measurement variability — giving referring clinicians calibrated bounds, not just numbers.