Overview

Schema Version

https://atlas.rsna.org/schemas/2025-11/model.json

Name

Deep Learning Tool for Automatic Measurement of Femoral Component Subsidence after THA

Link

https://doi.org/10.1148/ryai.210206

Indexing

Keywords: Femoral component subsidence, Total hip arthroplasty, Deep learning, U-Net, EfficientNetB0, Hip radiograph, Semantic segmentation

Content: MK

RadLex: RID35425, RID4619, RID5537

Author(s)

Pouria Rouzrokh

Cody C. Wyles

Shyam J. Kurian

Taghi Ramazanian

Jason C. Cai

Qiao Huang

Kuan Zhang

Michael J. Taunton

Hilal Maradit Kremers

Bradley J. Erickson

Organization(s)

Mayo Clinic

Version

1.0

License

URL: https://pubs.rsna.org/doi/10.1148/ryai.210206

Contact

Hilal Maradit Kremers, email: dev@null

Funding

Mayo Foundation Presidential Fund; National Institutes of Health grants R01AR73147 and P30AR76312.

Ethical review

Institutional review board approved; HIPAA-compliant; waiver of informed consent.

Date

Updated: 2022-05-04

Published: 2022-05-04

Created: 2021-07-25

References

[1] Rouzrokh P, Wyles CC, Kurian SJ, Ramazanian T, Cai JC, Huang Q, Zhang K, Taunton MJ, Maradit Kremers H, Erickson BJ. "Deep Learning for Radiographic Measurement of Femoral Component Subsidence Following Total Hip Arthroplasty". Radiology: Artificial Intelligence. 2022;4(3):e210206. 2022-05-04. doi:10.1148/ryai.210206. PMID: 35652119. PMCID: PMC9152683.

Model

Architecture

Dynamic U-Net semantic segmentation model with EfficientNetB0 encoder (ImageNet-pretrained) and squeeze-and-excitation attention in the decoder; image processing pipeline for landmark detection and subsidence calculation.

Availability

A stand-alone graphical user interface (GUI) was developed to run the tool; source framework used: segmentation_models.pytorch (https://github.com/qubvel/segmentation_models.pytorch).

Clinical benefit

Automates and standardizes measurement of femoral component subsidence to aid postoperative surveillance and potentially enable earlier detection of impending component failure.

Clinical workflow phase

Clinical decision support systems; workflow optimization for postoperative imaging assessment.

Degree of automation

Fully automatic; no user input or manual annotations required.

Indications for use

Quantification of femoral component subsidence between two serial AP hip radiographs in patients after total hip arthroplasty; intended for use in clinical and research settings for postoperative surveillance.

Input

Two serial AP hip radiographs (DICOM), 2048×2048 pixels, containing femur, femoral implant, and magnification markers.

Instructions

Upload/select two serial AP hip radiographs from the same patient; the tool automatically segments femur, implant, and magnification markers, identifies reference points, standardizes for magnification and femoral axis, and reports subsidence in millimeters. Standard AP projections with visible magnification markers are expected.

Limitations

Validated only on polished tapered cemented femoral stems; generalizability to other stem types not established. Performance may degrade with nonstandard AP radiographs, lower resolution images, missing magnification markers or markers of substantially different shape, unusual hardware, fractures, or severe heterotopic ossification; such cases led to segmentation failures and larger errors in a small number of patients.

Output

CDEs: RDE299, RDE311, RDE2090, RDE2514

Description: Quantitative estimate of femoral component subsidence (mm) between two serial radiographs; intermediate outputs include segmentation masks for femur, implant, and magnification markers and annotated reference lines/points.

Recommendation

Use as an automated measurement tool to support clinical and research assessment of stem subsidence after THA; larger, more heterogeneous validation is recommended before broad deployment across stem types.

Regulatory information

Authorization status: No regulatory clearance reported.

Sustainability

Runs in seconds on a stand-alone GUI; no special deep learning hardware required for the deployed GUI.

Use

Intended: Detection

Out-of-scope: Other

Excluded: Other

User

Intended: Radiologist, Subspecialist diagnostic radiologist, Researcher

Out-of-scope: Layperson

Excluded: Other