Overview

Schema Version

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

Name

Fully Automated ACL Tear Detection System for ACL Tears on Knee MRI

Link

https://dx.doi.org/10.1148/ryai.2019180091

Indexing

Keywords: ACL tear, knee MRI, deep learning, DenseNet, YOLO, LeNet, classification CNN, ligament isolation

Content: MR, MK

RadLex: RID48251, RID49389

SNOMED: 125601008

Author(s)

Fang Liu

Bochen Guan

Zhaoye Zhou

Alexey Samsonov

Humberto Rosas

Kevin Lian

Ruchi Sharma

Andrew Kanarek

John Kim

Ali Guermazi

Richard Kijowski

Organization(s)

Department of Radiology, University of Wisconsin School of Medicine and Public Health, Madison, WI

Department of Electrical and Computer Engineering, University of Wisconsin School of Engineering, Madison, WI

Department of Biomedical Engineering, University of Minnesota, Minneapolis, MN

Department of Radiology, Boston University School of Medicine, Boston, MA

Version

1.0

License

Text: © Radiological Society of North America, 2019

Contact

Fang Liu; email: ude.csiw@73uilf

Funding

National Institute of Arthritis and Musculoskeletal and Skin Diseases (R01-AR068373); National Institutes of Health (R01EB027087).

Ethical review

Retrospective study performed with institutional review board approval, HIPAA compliance, and waiver of informed consent.

Date

Updated: 2019-04-04

Published: 2019-05-08

Created: 2018-12-16

References

[1] Liu F, Guan B, Zhou Z, Samsonov A, Rosas H, Lian K, Sharma R, Kanarek A, Kim J, Guermazi A, Kijowski R. "Fully Automated Diagnosis of Anterior Cruciate Ligament Tears on Knee MR Images by Using Deep Learning". Radiology: Artificial Intelligence. 2019;1(3):180091.. 2019-05-08. doi:10.1148/ryai.2019180091. PMID: 32076658. PMCID: PMC6542618.

Model

Architecture

Cascaded deep CNNs: (1) section-selection CNN adapted from LeNet-5; (2) ligament isolation CNN adapted from YOLO to predict ACL-containing ROI; (3) classification CNN adapted from DenseNet with three dense blocks and global average pooling producing probability of ACL tear and diagnosis probability maps.

Clinical benefit

Automated detection of full-thickness ACL tears on knee MRI with performance comparable to clinical radiologists in a retrospective hold-out test set.

Decision threshold

Optimal operating point chosen by Youden index on validation/ROC analysis.

Degree of automation

Fully automated end-to-end processing at inference using three pre-trained CNNs.

Indications for use

Research setting: detection of presence or absence of full-thickness anterior cruciate ligament tear on sagittal PD-weighted and fat-suppressed T2-weighted fast spin-echo knee MRI.

Input

Sagittal proton density–weighted FSE and sagittal fat-suppressed T2-weighted FSE knee MR images; sections containing ACL are automatically selected and cropped to the intercondylar notch region prior to classification.

Instructions

Uses only sagittal PD-weighted and sagittal fat-suppressed T2-weighted FSE images. Pipeline automatically selects ACL-containing slices, isolates ACL region, and outputs tear probability and probability maps.

Limitations

Three CNNs trained separately (not single end-to-end training); relatively small training dataset; all images from a single institution, single 3.0-T scanner and protocol; verification bias since all subjects underwent arthroscopy; only full-thickness ACL tears included (no partial tears/intrasubstance sprains); generalizability to other scanners/protocols and to partial-thickness tears not established; combined radiologist+machine performance not evaluated.

Output

CDEs: RDE256.2, RDE231, RDE229

Description: Case-level probability and classification of presence/absence of ACL tear; per-slice and pixel-wise diagnosis probability maps highlighting regions contributing to the decision.

Recommendation

Research use; further technical development and multi-institutional prospective validation recommended before clinical deployment.

Regulatory information

Sustainability

Training times: section-selection 0.82 h; ligament isolation 5.11 h; classification 5.70 h on a workstation with Intel i7-7700K CPU, 32 GB RAM, dual Nvidia GTX 1080 GPUs. Average inference time ~9 seconds per subject.

Use

Intended: Automated detection of ACL tears on knee MRI (sagittal PD-wei...

Out-of-scope: General musculoskeletal MRI outside knee ACL assessment, Use with imaging sequences not included in training

Excluded: Detection of partial-thickness ACL tears/sprains (not represe...

User

Intended: Radiologists, Musculoskeletal radiologists, Researchers in medical imaging AI

Out-of-scope: Laypersons

Excluded: Patients