💻 computer science

Automated ACL Footprint Identification Using 3D Deep Learning

This study demonstrates that 3D deep learning models, particularly image-based architectures, can accurately identify the ACL femoral footprint center directly from 3D MR images, offering a promising clinical tool to improve tunnel placement and reduce reconstruction failures.

Original authors: Ruida Cheng, Ali Uneri, Gabriel Gibson, Frances T. Sheehan, Barry Boden

Published 2026-08-19

📖 4 min read☕ Coffee break read

CC BY 4.0

Original authors: Ruida Cheng, Ali Uneri, Gabriel Gibson, Frances T. Sheehan, Barry Boden

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Every year in the United States, hundreds of thousands of people undergo surgery to repair a torn anterior cruciate ligament, the tough band of tissue that stabilizes the knee. While the procedure is common, it does not always succeed. A frequent cause of failure is not the surgery itself, but the placement of the new ligament. Surgeons must drill a small tunnel into the thigh bone to anchor the new tissue, and if this tunnel is not positioned exactly where the original ligament once attached, the knee may remain unstable or develop arthritis later in life. For decades, finding this precise attachment point, known as the footprint, has relied on a surgeon's eye and experience, a manual process that leaves room for human error.

Recently, the field of medical imaging has begun to embrace artificial intelligence, using computer systems that can learn to recognize patterns in pictures. Most of this progress has focused on simply spotting whether a ligament is torn or healthy. However, identifying the exact center of the attachment point on the bone has remained a difficult challenge for these computers. A new study by researchers from the National Institutes of Health and Johns Hopkins University explores whether advanced three-dimensional deep learning can solve this specific problem. By teaching computers to analyze three-dimensional magnetic resonance images, the team aimed to create a tool that could automatically pinpoint the ideal spot for a surgeon to drill, potentially making knee reconstruction more reliable and consistent.

The researchers approached this challenge in two different ways, testing two distinct methods to see which could find the target location most accurately. The first method treated the bone as a digital 3D model, a mesh made of thousands of tiny points connected by lines, similar to a wireframe sculpture. They trained a computer system to look at the shape and surface of this digital bone to guess where the ligament attaches. The second method took a more direct approach, feeding the computer the raw three-dimensional magnetic resonance images themselves. This system learned to scan the volume of the knee, looking for the specific visual clues within the scan that indicate the footprint's center, much like a human radiologist would, but with the speed and precision of a machine.

To test these systems, the team used a massive collection of knee scans from a public database, comprising nearly 8,000 images from both left and right knees. They split this data, using the majority to teach the computers how to find the spot and saving a smaller portion to test their skills on new, unseen cases. The results showed that both methods worked, but one was clearly superior. The system that analyzed the raw three-dimensional images outperformed the one that relied on the digital 3D bone model. On average, the image-based system located the center of the attachment point with an error of just 2.1 millimeters, while the model-based system had an average error of 2.8 millimeters. In the world of knee surgery, where the difference between success and failure can be a matter of a few millimeters, this level of precision is significant.

The researchers also looked at how often the computer's guess fell within a safe range of the true location. For the image-based system, nearly 96 percent of the predictions for left knees and 93 percent for right knees were within 5 millimeters of the actual spot. To understand what this means for a patient, imagine the attachment area as a small circle on the bone. If a surgeon drills a hole that is slightly off-center, the new ligament might not cover the entire area, leading to instability. The study calculated that even with the small errors found in the computer's predictions, a standard surgical drill hole would still land mostly within the correct area. Only when using the largest possible drill size did a tiny fraction of the hole fall outside the ideal zone, suggesting that these computer tools could guide surgeons with a high degree of safety.

While the image-based method proved more accurate, the researchers noted that the 3D model approach still holds value. Because the model relies on the shape of the bone rather than the appearance of the ligament, it might work better in cases where the ligament is already torn and no longer visible in the scan. The study concludes that these artificial intelligence tools offer a feasible way to improve the accuracy of knee reconstruction surgery. By providing a precise, automated target for surgeons, this technology could help reduce the number of failed surgeries and improve the long-term health of the knee joint, turning a complex manual task into a guided, data-driven procedure.

Technical Summary: Automated ACL Footprint Identification Using 3D Deep Learning

Problem Statement
Anterior cruciate ligament (ACL) reconstruction failure is frequently attributed to femoral tunnel malpositioning, specifically the misalignment of the tunnel with the native ACL femoral footprint. Such inaccuracies can lead to joint instability, meniscal pathology, and osteoarthritis. While current artificial intelligence (AI) research in orthopedics focuses primarily on ACL segmentation and rupture classification using magnetic resonance (MR) images, the automated identification of the ACL femoral footprint center remains underexplored. Existing methods for footprint localization are largely manual, and no prior studies have thoroughly investigated the accuracy of automatically locating the ACL femoral footprint center using deep learning.

Methodology
The study proposes two distinct 3D deep learning architectures to automate ACL femoral footprint center identification, treating the task as a landmark detection problem to address the spatial imbalance of a single target point within a large volumetric domain.

3D Mesh-Based Geometric Model (DGCNN):
- Input: 3D femoral surface meshes (PLY format) and annotated landmark points.
- Architecture: Built upon Dynamic Graph Convolutional Neural Networks (DGCNN).
- Process:
  - Preprocessing: Deterministic sampling reduces vertex count while maintaining uniform spacing. Vertices are represented by 7-dimensional feature vectors (coordinates, surface normals, and distance from center).
  - Feature Extraction: Edge convolution blocks dynamically recompute k-nearest neighbors to capture local geometric shapes and multiscale representations.
  - Prediction: A multilayer perceptron generates a vertex-wise probability distribution (heatmap) for the center and surrounding sampling points.
  - Loss Function: Combines Gaussian heatmap supervision with coordinate regression loss, weighting points closer to the true center higher.
3D Volumetric Image-Based Model (Enhanced nnLandmark):
- Input: 3D MR image volumes (Axial 3D DESS sequences).
- Architecture: An extension of the nnLandmark architecture, based on a 3D residual-based nnUNet backbone.
- Enhancements:
  - Squeeze-and-Excitation (SE) Attention: Integrated into deeper encoder stages and the bottleneck to recalibrate feature maps and suppress noise.
  - Multi-scale High-Resolution Fusion: Intermediate encoder features are fused with the decoder output via trilinear interpolation to recover high-resolution spatial details necessary for small targets.
  - Top-K Peak Decoding: Inference involves converting logits to probability heatmaps, filtering by a threshold, and selecting the top $K$ peaks (default $K=5$ ) based on spatial local maxima to ensure stable single-voxel predictions.

Dataset and Training
The models were trained and tested on 7,969 3D knee MR images from the publicly available 3DReasonKnee dataset (derived from the Osteoarthritis Initiative). The dataset included 4,883 right and 3,087 left knee image sets.

Data Split: 80% for training, 20% for testing.
Ground Truth Generation: Since the source dataset lacked footprint landmarks, the authors derived ACL footprint centers and 3D femoral meshes from existing ACL and femur segmentation masks.
Mesh Specifics: Each mesh case included one center point and 20 surrounding sampling points (21 landmarks total).

Results
Both models demonstrated feasibility, but the image-based approach outperformed the mesh-based approach.

3D Mesh-Based Model (DGCNN):
- Right Femur: Mean error of 2.638 mm; Median error of 2.096 mm.
- Left Femur: Mean error of 3.170 mm; Median error of 2.378 mm.
- PCK (Percentage of Correct Keypoints): 47.6% (Right) and 44.8% (Left) of predictions were within 2 mm of the ground truth.
3D Image-Based Model (Enhanced nnLandmark):
- Right Femur: Mean error of 2.353 mm; Median error of 1.600 mm.
- Left Femur: Mean error of 1.816 mm; Median error of 1.275 mm.
- PCK: 68.7% (Right) and 77.8% (Left) of predictions were within 2 mm of the ground truth.
- P95 Error: 95% of predictions fell within 6.82 mm (Right) and 4.60 mm (Left) of the ground truth.

Significance and Claims
The authors claim this study provides a feasible clinical approach for automated ACL footprint localization. Key claims include:

Superiority of Image-Based Methods: The 3D volumetric image-based model achieved lower average errors (2.1 mm overall) compared to the mesh-based model (2.8 mm overall).
Surgical Relevance: Based on a geometric analysis assuming a circular footprint of 126.8 mm², the average errors of both models suggest that a standard 7 mm drill hole would remain entirely within the native footprint. Even with a 10 mm drill hole, only 6% (image-based) and 14% (mesh-based) of the hole would fall outside the footprint.
Applicability to Torn ACLs: While the study utilized images with intact ACLs (where the footprint is visible), the authors suggest the mesh-based model holds specific promise for cases with torn ACLs where the footprint is not directly visible on MR images, as it relies on femoral geometry rather than ligament visibility.
Novelty: The authors state that, to their knowledge, this is the first study to evaluate the accuracy of automatically locating the ACL femoral footprint center, filling a gap between automated image analysis and surgical planning.

The paper concludes that 3D deep learning can improve the accuracy of tunnel placement in ACL reconstruction, potentially reducing graft failure and post-surgical complications, though further work is needed to validate generalizability across torn ACLs and reduce error margins further.

Technical Summary: Automated ACL Footprint Identification Using 3D Deep Learning

More like this