Reconstructing the Distal Femoral Surface from Knee MRI Using a Continuous Implicit Surface Representation for Patient-Specific Implant Design

Automating patient-specific implant (PSI) design remains an open challenge in orthopedics. Most deep learning methods predict binary voxel masks, from which reconstructed surfaces often exhibit staircase artifacts that require manual repair before use in computer-aided design (CAD). To address this limitation, we present Ortho-Synth, a continuous implicit surface reconstruction framework that predicts the distal femoral surface from knee MRI as a basis for PSI design. Rather than predicting a discrete voxel label at each location, Ortho-Synth regresses a continuous signed distance field (SDF) directly from the 3D MRI volume. The distal femoral surface is then extracted from the predicted SDF and smoothed using a volume-preserving Taubin filter. The framework was developed using 352 knee magnetic resonance imaging (MRI) series from 216 participants in the Osteoarthritis Initiative and evaluated on all 71 held-out series. An anatomical-plausibility check, applied identically to all methods and to the automatically generated reference masks, was satisfied by the same 29 series (40.8%). The remaining 42 series (59.2%) were coronal 3D fast low-angle shot (FLASH) acquisitions for which the reference masks themselves comprised thin slabs. Among the 29 evaluable series, Ortho-Synth achieved a mean Dice score of 0.881 ± 0.023 and a 95th-percentile Hausdorff distance (HD95) of 5.865 ± 1.007 voxels. After identical boundary-closing and Taubin-smoothing refinement, nnU-Net achieved higher overlap (Dice score, 0.960 ± 0.014) and lower surface distances, whereas Ortho-Synth produced fewer connected components (i.e., separate mesh pieces; 66.6 vs. 336.8), lower surface roughness (11.4° vs. 22.2°), and a higher proportion of watertight, edge-manifold meshes (86.2% vs. 0%) (all p < 0.001). A paired ablation study comparing SDF regression with a binary-voxel network using the same architecture and training protocol revealed a similar trade-off: SDF regression reduced overlap (Dice score, 0.882 vs. 0.931) but markedly reduced fragmentation and surface roughness while increasing watertightness. These findings suggest that SDF-based reconstruction may reduce the manual surface repair required before CAD modeling, at the cost of lower voxel-level overlap with the reference masks. However, because the unvalidated Otsu-derived reference masks included non-femoral structures, the reported agreement metrics do not establish anatomical accuracy, while series-level data partitioning compromises the independence of the test set.

Authors

Institutions

Publication Details

Journal
AI
Published
2026-11-18
DOI
https://doi.org/10.3390/ai7110491
Primary Topic
Medical Image Segmentation Techniques
Type
article
Field-Weighted Citation Impact
0.00
Controls
|||
ALL TIME
JAN
FEB
MAR
APR
MAY
JUN
JUL
AUG
SEP
OCT
NOV
article

Reconstructing the Distal Femoral Surface from Knee MRI Using a Continuous Implicit Surface Representation for Patient-Specific Implant Design

Yen-Ching Chang, Ji-Ci Li
3 links
AI
Medical Image Segmentation Techniques
article

Reconstructing the Distal Femoral Surface from Knee MRI Using a Continuous Implicit Surface Representation for Patient-Specific Implant Design

Yen-Ching Chang, Ji-Ci Li
article en

Abstract

Automating patient-specific implant (PSI) design remains an open challenge in orthopedics. Most deep learning methods predict binary voxel masks, from which reconstructed surfaces often exhibit staircase artifacts that require manual repair before use in computer-aided design (CAD). To address this limitation, we present Ortho-Synth, a continuous implicit surface reconstruction framework that predicts the distal femoral surface from knee MRI as a basis for PSI design. Rather than predicting a discrete voxel label at each location, Ortho-Synth regresses a continuous signed distance field (SDF) directly from the 3D MRI volume. The distal femoral surface is then extracted from the predicted SDF and smoothed using a volume-preserving Taubin filter. The framework was developed using 352 knee magnetic resonance imaging (MRI) series from 216 participants in the Osteoarthritis Initiative and evaluated on all 71 held-out series. An anatomical-plausibility check, applied identically to all methods and to the automatically generated reference masks, was satisfied by the same 29 series (40.8%). The remaining 42 series (59.2%) were coronal 3D fast low-angle shot (FLASH) acquisitions for which the reference masks themselves comprised thin slabs. Among the 29 evaluable series, Ortho-Synth achieved a mean Dice score of 0.881 ± 0.023 and a 95th-percentile Hausdorff distance (HD95) of 5.865 ± 1.007 voxels. After identical boundary-closing and Taubin-smoothing refinement, nnU-Net achieved higher overlap (Dice score, 0.960 ± 0.014) and lower surface distances, whereas Ortho-Synth produced fewer connected components (i.e., separate mesh pieces; 66.6 vs. 336.8), lower surface roughness (11.4° vs. 22.2°), and a higher proportion of watertight, edge-manifold meshes (86.2% vs. 0%) (all p < 0.001). A paired ablation study comparing SDF regression with a binary-voxel network using the same architecture and training protocol revealed a similar trade-off: SDF regression reduced overlap (Dice score, 0.882 vs. 0.931) but markedly reduced fragmentation and surface roughness while increasing watertightness. These findings suggest that SDF-based reconstruction may reduce the manual surface repair required before CAD modeling, at the cost of lower voxel-level overlap with the reference masks. However, because the unvalidated Otsu-derived reference masks included non-femoral structures, the reported agreement metrics do not establish anatomical accuracy, while series-level data partitioning compromises the independence of the test set.

AIVol. 7(11)
Chung Shan Medical University Hospital (TW), Chung Shan Medical University (TW)
Good health and well-being
Medical Image Segmentation Techniques
AI Navigator

Ask Laika to Summarize, Analyze, and Connect papers live on the map.

Summarize Papers & Methodologies

Extract key findings, datasets, and comparative methods across publications.

Benchmark Rankings & Visual Analytics

Rank top research institutions, authors, funders, topics, and journals by Field-Weighted Citation Impact (FWCI) and paper volume with instant charts.

Connect Distant Disciplines

Bridge topological clusters on the map to find hidden collaborative intersections.

Laika access coming soon