A multimodal deep learning radiomics model for predicting degenerative meniscus tear after arthroscopy

Area of Science:

• Orthopedics
• Radiology
• Artificial Intelligence

Background:

• Degenerative meniscus tears often coexist with osteoarthritis, complicating prognosis after arthroscopic partial meniscectomy (APM).
• Predicting patient outcomes following APM is challenging due to these comorbidities.

Purpose of the Study:

• To develop and validate a multimodal deep learning radiology (MDLR) model.
• To integrate deep learning radiology (DLR) scores from MRI with clinical variables for improved prognostic accuracy.

Main Methods:

• Retrospective enrollment of 452 patients undergoing APM for degenerative meniscus tears (February 2020 - February 2022).
• Extraction of DLR features from knee MRI scans.
• Development of an MDLR model incorporating DLR signatures and clinical variables for prognostic risk stratification.

Main Results:

• A standalone DLR model showed limited predictive performance (ROC curves ranging from 0.720-0.780).
• Multivariate analysis identified key prognostic indicators: gender, height, weight, pain duration, ESR, and VAS.
• The integrated MDLR model, particularly using Light Gradient Boosting Machine, significantly improved performance (ROC curves up to 0.951) and stratified patients into four prognostic levels.

Conclusions:

• The developed MDLR tool effectively stratifies prognosis for APM in degenerative meniscus tear patients.
• Pain exacerbation time is a crucial prognostic factor identified by the MDLR model.
• MDLR demonstrates significant potential to assist physicians in therapeutic decision-making and surveillance strategies.