From 2D Shadows to 3D Precision: How a UCL Team is Revolutionizing Hip Replacement Planning

Total Hip Arthroplasty (THA) is one of the most successful surgeries in modern medicine, yet it still faces a persistent “Goldilocks” problem: finding the perfect fit for the acetabular cup.

At MICCAI 2025, a team from University College London (UCL) presented an oral presentation that tackles a major clinical pain point. They’ve developed a way to reconstruct a patient’s 3D acetabular surface with extreme precision — using only a few standard X-rays instead of a high-radiation CT scan.

The Clinical Dilemma: Accuracy vs. Radiation

Currently, most surgeons rely on 2D template matching on X-ray films to choose the implant size. The problem? It only works about 70% of the time. Even veteran surgeons struggle with magnification distortion and variations in patient positioning.

The stakes of a “misfit” are high:

• Too large: Risk of bone fractures or soft tissue impingement.
• Too small: Risk of implant loosening, premature failure, and increased infection risk due to longer revision surgeries.

While 3D CT-based planning offers over 99% accuracy, it’s rarely the standard. Following the ALARA (As Low As Reasonably Achievable) principle, doctors avoid unnecessary radiation. A pelvic CT delivers roughly 5–10 mSv, whereas a single X-ray is only about 0.6 mSv. Even taking three X-rays keeps the radiation dose significantly lower than a single CT scan.

Moving Beyond “Black Box” Deep Learning

Technically, this paper takes a refreshing detour from the current trend of “brute force” deep learning or standard Statistical Shape Models (SSM). Because the acetabulum is anatomically complex and often obscured in projections, standard models often fail to resolve “projection ambiguity.”

Instead, the UCL team proposed a framework combining SRVF Elastic Registration and ED Graph optimization.

1. SRVF (Square Root Velocity Function) Elastic Registration

Rooted in shape analysis theory from around 2010, SRVF changes the game. While traditional ICP (Iterative Closest Point) algorithms treat edges as a collection of discrete points, SRVF models 2D contours as continuous lines. By treating geometric alignment as a mathematical function-matching problem, the system is far less likely to get “stuck” in local optima.

2. ED (Embedded Deformation) Graph Optimization

Borrowing a powerful concept from computer graphics, the team uses a standard hemispherical model as an initial template. This model is driven by a deformation graph that undergoes non-rigid warping. Through non-linear least squares optimization, the algorithm ensures that the 3D model’s projected silhouette perfectly matches the real-world observations on the X-ray.

The Results: Clinical-Grade Precision

The framework was validated on 5 clinical cases using just three preoperative X-rays (one AP view and two lateral views). The results are impressive:

• Reconstruction Accuracy: The Mean Absolute Error (MAE) was kept between 0.98mm and 1.63mm.
• The “Sizing” Test: Since standard acetabular cups typically come in 4mm size increments, an error margin of ~1mm is a “safe zone.”

In every single test case, the algorithm’s predicted size matched the actual size chosen by the surgeons during the operation.

Final Thoughts

By bridging the gap between the low cost of X-rays and the high precision of 3D modeling, this UCL study offers a glimpse into a future where “personalized surgery” doesn’t require high radiation or expensive equipment. It’s a masterclass in applying classic geometric theories to modern clinical hurdles.

| Find papers faster on arXivSub with AI summary (CVPR/ICCV/ICML/ICLR/NeurIPS/AAAI/MICCAI)