2025 ISAKOS Congress in Munich, Germany

2025 ISAKOS Biennial Congress ePoster

 

Diagnosis Of Medial Meniscus Posterior Root Tear On Magnetic Resonance Imaging Using Deep Learning

Yuta Nakanishi, MD, PhD, Kobe, Hyogo JAPAN
Takehiko Matsushita, MD, PhD, Kobe, Hyogo JAPAN
Atsuyuki Inui, MD, PhD, Kobe JAPAN
Kyohei Nishida , MD, PhD, Kobe, Hyogo JAPAN
Kanto Nagai, MD, PhD, Kobe, Hyogo JAPAN
Yuichi Hoshino, MD, PhD, Kobe, Hyogo JAPAN
Ryosuke Kuroda, MD, PhD, Kobe, Hyogo JAPAN

Kobe University Graduate School of Medicine, Kobe, Hyogo, JAPAN

FDA Status Not Applicable

Summary

Diagnosis of medial meniscus posterior root tear (MMPRT) on magnetic resonance imaging using deep learning is accurate and precise based on the current model and may potentially be used in assisting diagnosis of MMPRT with high consistency regardless of specialty of the physician.

ePosters will be available shortly before Congress

Abstract

Introduction

Medial meniscus posterior root tear (MMPRT) leads to extrusion of the medial meniscus, ultimately contributing to the progression of knee osteoarthritis. Accurate and timely diagnosis of MMPRT, followed by appropriate treatment is therefore crucial. While MMPRT can be detected via magnetic resonance imaging (MRI), the associated signal changes can be subtle, potentially resulting in misdiagnosis or delayed diagnosis and treatment. Therefore, widespread use of technology that aids in the consistent diagnosis of MMPRT, regardless of the physician's specialty, could be highly beneficial. This study aims to evaluate the accuracy and precision of diagnosing MMPRT on MRI using deep learning technology. We hypothesized that deep learning would enable highly accurate and precise MMPRT diagnosis on MRI.

Methods

Preoperative coronal and sagittal T2 STIR and fat suppression MRI images were obtained from picture archiving and communication system (PACS) data of patients that underwent arthroscopic knee procedures. MRI images for MMPRT cases were obtained from patients with MMPRT confirmed by arthroscopy upon repair of MMPRT or other arthroscopic procedures. Non-MMPRT images were attained from patients who underwent arthroscopic procedures and were confirmed to not have MMPRT. 100 coronal and 100 sagittal images for the MMPRT group, and 100 coronal and 100 sagittal images for the non-MMPRT group were included. 80% (80 coronal and 80 sagittal images) of images from each group were randomly selected as training data, and the remaining (20 coronal and 20 sagittal images) were used for test data. Transfer learning using efficientNet was conducted. Model evaluation was performed using confusion matrix and ROC curve. To visualize important features, occlusion sensitivity was used.

Results

MMPRT images from 22 patients (7 males, 15 females; mean 60.2 years) and non-MMPRT images from 23 patients (12 males, 11 females; mean 48.7 years) were included. The model had an accuracy of 0.86, precision of 0.91. The area under the curve (AUC) based on receiver operating characteristic (ROC) was 0.90. Using occlusion sensitivity to visualize the region of interests, the deep learning model focused on the region between posterior horn and root of the medial meniscus.

Conclusion

Diagnosis of medial meniscus posterior root tear (MMPRT) on magnetic resonance imaging using deep learning based on the current model used is accurate. Increasing the number of learned images and further enhancement of the model may increase the accuracy and precision of diagnosis.