2608.07368-Measurements Automatically Extracted from Zero Echo Time MRI Using Deep Learning Image Segmentation and Geometric Modeling Agree with Expert Manual Readings

Introduction to the show: ident: Paper Radio. Generated commentary on the latest Artificial Intelligence papers.

Tom: Next we'll be talking about the paper "Measurements Automatically Extracted from Zero Echo Time MRI Using Deep Learning Image Segmentation and Geometric Modeling Agree with Expert Manual Readings".

Jane: The paper was written by Jack Consolini, Eric A. Bogner, Meghan Sahr, Matthew F. Koff, Kevin M. Koch et al. from Hospital for Special Surgery.

Tom: Stay tuned as we take you through the paper and discuss its implications.

Title: Tom: This paper comes out of the Hospital for Special Surgery in New York, and the first author is Jack Consolini, a biomedical engineer there. The title really does summarize the whole project — the team pulled hip measurements automatically out of a special kind of MRI, then showed those numbers agree with readings made by expert radiologists by hand.

Jane: Let's unpack the special MRI first, because that's what makes everything else possible. It's called zero echo time, or ZTE, and it makes cortical bone show up bright on an MRI scan, almost like a CT image, but without any ionizing radiation. That's a real advantage because these precise bone angles usually mean sending a young patient for a CT scan.

Tom: Right, and the angles matter for a condition called femoroacetabular impingement, FAI, where the ball of the hip joint and the socket don't fit together smoothly, so bone hits bone during movement and slowly damages the joint. Hip dysplasia is the related problem where the socket is too shallow. Both are major causes of early arthritis in young, active people.

Jane: And that's the population this really touches — mostly athletes and women of reproductive age. Today, a complete workup can involve a CT for the three dee bone geometry plus an MRI for the soft tissue like the labrum and cartilage. This paper asks whether one MRI can cover both.

Tom: They had good reason to think it's possible, because a study from the same group, by Breighner and colleagues, had already shown that trained readers can measure these angles on ZTE MRI and agree with CT. The new step is taking the human out of the measurement loop entirely.

Jane: And that's where deep learning comes in. They train a neural network to find the femur, the pelvis, and specific bone landmarks, and then a geometric algorithm computes eight angles from those segmentations. If it works, a single radiation-free scan could give a surgeon everything needed to plan a hip preservation operation.

Tom: The appeal is obvious. The real question is whether the automated measurements actually hold up against the experts.

Jane: And that's exactly how the study is designed — with a held-out test set the model never trained on. Let's look at how they built it and what they found.

Summary: Tom: So the segmentation engine is nnU-Net, which is a self-configuring neural network that many medical imaging labs use as their default tool. The team trained it on a hundred manually curated hips, segmenting the femur and pelvis plus three small landmark spheres placed at the lateral acetabulum, the medial weight-bearing acetabulum, and the distal greater trochanter.

Jane: The landmarks are the clever part, because they capture where a radiologist would actually look when measuring. On the thirty-five held-out test hips, the bone segmentation was excellent — Dice scores of 0 point 98 for the femur and 0 point 97 for the pelvis. The landmarks were weaker, with Dice around 0 point 65 to 0 point 83, but their center-point errors were still small, under a millimeter for the femoral head.

Tom: And then comes the real test — the automated angles compared against the mean of two experienced musculoskeletal radiologists, one with twenty years of experience and one with ten. For acetabular version at three clock positions, the coronal center-edge angle, and the Tönnis angle, the ICCs came out between 0 point 92 and 0 point 96, which is excellent agreement.

Jane: But the alpha angle and femoral neck-shaft angle lagged, with fair agreement around 0 point 45 and 0 point 55. The mid-acetabular sagittal center-edge score was good, at 0 point 74. What's striking, though, is how the two human experts performed on exactly those same angles.

Tom: Their interrater ICC for alpha and femoral neck-shaft was just 0 point 20. On the sagittal center-edge they only managed 0 point 45, and they had a big systematic disagreement of nearly eight degrees on that one. One radiologist re-read ten hips, and the alpha angle ICC came out negative.

Jane: So the model struggles precisely where the experts struggle.

Tom: Exactly. And the Bland–Altman numbers support that framing. For most angles, the model's limits of agreement were narrower than the limits between the two raters. The alpha angle had the widest spread all around, but the model's spread of about 8 point 3 degrees was actually tighter than the raters' 11 point 5 degrees.

Jane: So the authors argue the automated method is reproducible, even when it doesn't numerically match a given reader's habit. That seems like a fair way to read the results.

Tom: It is. And it sets up the clinical question — what does this actually change for patients and for the people reading their scans?

Improvements: Tom: The main clinical suggestion is that automated ZTE morphometry could let surgeons skip the adjunct CT in the pre-operative workup for hip preservation. Right now a patient often gets an MRI for the soft tissue and then a separate CT for the bone angles, and this pipeline could let a single exam serve both purposes.

Jane: And it's not only about avoiding radiation, though that's enormous for young athletes and women of reproductive age. It's also about standardization. Manual measurement takes a radiologist's time and carries observer bias, whereas the automated pipeline applies the same geometric definitions to every hip, every time.

Tom: They also make a smart point about post-operative imaging. After hip preservation surgery, patients return for follow-ups, and comparing angles across visits is far more meaningful if the same automated method computes them each time, rather than different readers with slightly different habits.

Jane: Drift between readers is a real problem in follow-up imaging.

Tom: Right. And there's a longitudinal opportunity too. Because ZTE MRI carries no radiation, you could image a young athlete repeatedly across seasons and track whether the impingement morphology is progressing. You wouldn't do that with CT.

Jane: Now, they're honest about the limits. The study is single institution, single scanner vendor, one ZTE acquisition approach. The authors think the method transfers to other sequences as long as the femur and acetabulum can be segmented, but they haven't proven it.

Tom: And the training data has caveats — most landmark labels were placed by a trained research engineer under radiologist oversight rather than directly by the radiologists. The intra-rater reliability check used only ten hips, which is a small sample, even if it matches the earlier ZTE study from this group.

Jane: They also worry about severe deformity or imaging artifact degrading the landmarks, especially at the medial acetabulum, and they admit the alpha angle and sagittal center-edge depend on algorithmic choices that may not mirror manual convention.

Tom: The future work they sketch is very practical — testing whether automated angles increase referring physicians' confidence, and quantifying the time and cost savings of dropping CT from the workflow. Those are the questions that decide whether hospitals actually change their ordering habits.

Jane: Exactly. The technical validation is one thing, but adoption depends on clinicians trusting the numbers enough to skip the CT.

Tom: And that's where the abstract comes in, because it's the distilled version of the whole study that those clinicians will actually read. Let's look at the first page.

First Page: Tom: The abstract sets the frame from the first sentence. CT is the reference for three-dimensional bone measurement in FAI, but it delivers ionizing radiation and requires manual measurement. ZTE MRI visualizes cortical bone and manual readings agree with CT, yet automated extraction had remained limited.

Jane: Then the purpose is stated directly — to develop and validate automated Feye angle computation with good to excellent agreement against expert manual readings. And they set a clear hypothesis: the automated angles would agree with the experts at a level comparable to the agreement between two experts.

Tom: They describe the design as a cross-sectional study, level of evidence 3, and they did a prospective sample size calculation using Fisher's z-transformation. That calculation required at least 23 hips for 95 percent power, and they ended up with 35 test hips, so the validation is adequately powered.

Jane: The methods summary mentions the nnU-Net trained on a hundred curated hips and the eight angle measurements. The results quote bone Dice above 0 point 96, landmark Dice from 0 point 65 to 0 point 83, and median landmark errors from 0 point 38 millimeters on the femoral head to under 2 point 5 millimeters on the medial acetabulum and greater trochanter.

Tom: The headline result is the model's agreement with the rater mean — excellent for acetabular versions, coronal center-edge, and Tönnis, with ICCs from 0 point 92 to 0 point 96, good for sagittal center-edge at 0 point 74, and fair for alpha and femoral neck-shaft. They define the poor, fair, good, and excellent thresholds up front, so there's no ambiguity about what those labels mean.

Jane: What I appreciate is that they put the Bland–Altman comparison in the abstract too. The model's limits of agreement were narrower than the interrater limits for most angles, which directly addresses the worry that a machine can't match human judgment.

Tom: The conclusion is measured — fully automated morphometric assessment from ZTE MRI is feasible and performs comparably to an expert reader for most coverage and version angles. They don't stretch the claim to the alpha angle.

Jane: And the clinical relevance statement is the whole paper in one sentence. It says this approach may reduce adjunct CT for pre-operative morphometry in athletes, young active patients, and women of reproductive age, providing standardized automated measurements from a single radiation-free MRI examination.

Tom: Reading the abstract after going through the details, every number we dug into shows up there, no more and no less. It's an honest summary.

Jane: It is. And we've covered the full arc — the motivation, the methods, the results, and the clinical case. I think we're ready to wrap up.

Conclusion: Tom: So the summary is straightforward. This is a validation study from the Hospital for Special Surgery showing that a fully automated ZTE MRI pipeline can measure most hip angles as reliably as expert radiologists, with no radiation and no manual slice-by-slice work.

Jane: The strong results are the coverage and version angles — acetabular version at the three clock positions, coronal center-edge, and Tönnis — where the model reached excellent agreement with the rater mean and produced limits of agreement narrower than the experts achieved with each other.

Tom: And the alpha angle and femoral neck-shaft angle remain difficult, but the paper shows the experts struggle in the same places. The model's fair agreement makes sense in that context. The study itself was built on a solid cohort of 73 participants and 135 hips, with a held-out test set that cleared their prospective sample size requirement.

Jane: The bigger picture is a single MRI exam that gives bone geometry and soft tissue in one go, standardized across patients and visits, and safe to repeat in young athletes and women of reproductive age. That's a genuine shift from today's CT-plus-MRI workflow.

Tom: The next steps are multi-center validation and real-world workflow studies, measuring whether referring clinicians trust the automated numbers enough to skip the CT. That's the operational question that determines adoption.

Jane: For now, the paper makes a solid case that automation has reached parity with expert readers on the angles that matter most for hip preservation decisions. And that's a good place to leave it.

Tom: Agreed. Let's close this one out and move on to the next paper.