Objectives Knee injuries frequently require Magnetic Resonance Imaging (MRI) evaluation, increasing radiologists’ workload. This study evaluates the impact of a Knee AI assistant on radiologists’ diagnostic accuracy and efficiency in detecting anterior cruciate ligament (ACL), meniscus, cartilage, and medial collateral ligament (MCL) lesions on knee MRI exams. Materials and methods This retrospective reader study was conducted from January 2024 to April 2024. Knee MRI studies were evaluated with and without AI assistance by six radiologists with between 2 and 10 years of experience in musculoskeletal imaging in two sessions, 1 month apart. The AI algorithm was trained on 23,074 MRI studies separate from the study dataset and tested on various knee structures, including ACL, MCL, menisci, and cartilage. The reference standard was established by the consensus of three expert MSK radiologists. Statistical analysis included sensitivity, specificity, accuracy, and Fleiss’ Kappa. Results The study dataset involved 165 knee MRIs (89 males, 76 females; mean age, 42.3 ± 15.7 years). AI assistance improved sensitivity from 81% (134/165, 95% CI = [79.7, 83.3]) to 86%(142/165, 95% CI = [84.2, 87.5]) (p < 0.001), accuracy from 86% (142/165, 95% CI = [85.4, 86.9]) to 91%(150/165, 95% CI = [90.7, 92.1]) (p < 0.001), and specificity from 88% (145/165, 95% CI = [87.1, 88.5]) to 93% (153/165, 95% CI = [92.7, 93.8]) (p < 0.001). Sensitivity and accuracy improvements were observed across all knee structures with varied statistical significance ranging from < 0.001 to 0.28. The Fleiss’ Kappa values among readers increased from 54% (95% CI = [53.0, 55.3]) to 78% (95% CI = [76.6, 79.0]) (p < 0.001) post-AI integration. Conclusion The integration of AI improved diagnostic accuracy, efficiency, and inter-reader agreement in knee MRI interpretation, highlighting the value of this approach in clinical practice. Key Points <Emphasis Type="BoldItalic">Question</Emphasis> Can artificial intelligence (AI) assistance improve the diagnostic accuracy and efficiency of radiologists in detecting main lesions anterior cruciate ligament, meniscus, cartilage, and medial collateral ligament lesions in knee MRI? <Emphasis Type="BoldItalic">Findings</Emphasis> AI assistance in knee MRI interpretation increased radiologists’ sensitivity from 81 to 86% and accuracy from 86 to 91% for detecting knee lesions while improving inter-reader agreement (p < 0.001). <Emphasis Type="BoldItalic">Clinical relevance</Emphasis> AI-assisted knee MRI interpretation enhances diagnostic precision and consistency among radiologists, potentially leading to more accurate injury detection, improved patient outcomes, and reduced diagnostic variability in musculoskeletal imaging. Graphical Abstract

Impact of AI assistance on radiologist interpretation of knee MRI

期刊论文
发表日期：

2025年7月31日

查看全文

Guillaume Herpe,
Tom Vesoul,
Pascal Zille,
Etienne Pluot,
Raphael Guillin,
Benoit Rizk,
Roberto Ardon,
Chloe Adam,
Gaspard d’Assignies,
Pedro Augusto Gondim Teixeira

摘要

Objectives

Knee injuries frequently require Magnetic Resonance Imaging (MRI) evaluation, increasing radiologists’ workload. This study evaluates the impact of a Knee AI assistant on radiologists’ diagnostic accuracy and efficiency in detecting anterior cruciate ligament (ACL), meniscus, cartilage, and medial collateral ligament (MCL) lesions on knee MRI exams.

Materials and methods

This retrospective reader study was conducted from January 2024 to April 2024. Knee MRI studies were evaluated with and without AI assistance by six radiologists with between 2 and 10 years of experience in musculoskeletal imaging in two sessions, 1 month apart. The AI algorithm was trained on 23,074 MRI studies separate from the study dataset and tested on various knee structures, including ACL, MCL, menisci, and cartilage. The reference standard was established by the consensus of three expert MSK radiologists. Statistical analysis included sensitivity, specificity, accuracy, and Fleiss’ Kappa.

Results

The study dataset involved 165 knee MRIs (89 males, 76 females; mean age, 42.3 ± 15.7 years). AI assistance improved sensitivity from 81% (134/165, 95% CI = [79.7, 83.3]) to 86%(142/165, 95% CI = [84.2, 87.5]) (p < 0.001), accuracy from 86% (142/165, 95% CI = [85.4, 86.9]) to 91%(150/165, 95% CI = [90.7, 92.1]) (p < 0.001), and specificity from 88% (145/165, 95% CI = [87.1, 88.5]) to 93% (153/165, 95% CI = [92.7, 93.8]) (p < 0.001). Sensitivity and accuracy improvements were observed across all knee structures with varied statistical significance ranging from < 0.001 to 0.28. The Fleiss’ Kappa values among readers increased from 54% (95% CI = [53.0, 55.3]) to 78% (95% CI = [76.6, 79.0]) (p < 0.001) post-AI integration.

Conclusion

The integration of AI improved diagnostic accuracy, efficiency, and inter-reader agreement in knee MRI interpretation, highlighting the value of this approach in clinical practice.

Key Points

Question Can artificial intelligence (AI) assistance improve the diagnostic accuracy and efficiency of radiologists in detecting main lesions anterior cruciate ligament, meniscus, cartilage, and medial collateral ligament lesions in knee MRI?

Findings AI assistance in knee MRI interpretation increased radiologists’ sensitivity from 81 to 86% and accuracy from 86 to 91% for detecting knee lesions while improving inter-reader agreement (p < 0.001).

Clinical relevance AI-assisted knee MRI interpretation enhances diagnostic precision and consistency among radiologists, potentially leading to more accurate injury detection, improved patient outcomes, and reduced diagnostic variability in musculoskeletal imaging.

Graphical Abstract