Improved diagnosis of tibiofemoral cartilage defects on MRI images using deep learning

Abstract

Introduction: MRI is the modality of choice for cartilage imaging, however, its diagnostic performance is variable. Objectives: We aimed to evaluate whether deep learning can be utilized to accurately identify cartilage defects when applied to the interpretation of knee MRI images and compare deep learning's performance to that of an orthopaedic trainee and an orthopaedic surgeon. Methods: We analyzed data from patients who underwent knee MRI evaluation and subsequent arthroscopic knee surgery (207 with-, 90 without cartilage defect). Patients' arthroscopic findings were compared to preoperative MRI images to verify the presence or absence of isolated tibiofemoral cartilage defects. For each patient, the most representative MRI image slice of the patient's condition was selected (defect or no-defect) from a coronal- and sagittal view. We developed three convolutional neural networks (CNNs) to analyze the images: CNN-1 trained on the images of the sagittal and coronal views; CNN-2 trained on the images of the sagittal view; CNN-3 trained on the images of the coronal view. We implemented image-specific saliency maps to visualize the CNNs decision-making process. The same test dataset images were then provided to an experienced orthopaedic surgeon and an orthopaedic trainee. Results: Saliency maps demonstrated that the CNNs learned to focus on the clinically relevant areas of the MRI. The CNN-1 achieved higher performance (sensitivity-86.96, specificity-100%, positive predictive value [PPV]-100%, negative predictive value [NPV]-66.67%) than the orthopaedic surgeon (sensitivity-82.61%, specificity-83.33%, PPV- 95%, NPV-55.56%), Conclusions: CNN can be used to enhance the diagnostic performance of MRI in identifying isolated tibiofemoral cartilage defects.

Keywords

• Cartilage defect
• Knee
• Mri
• Deep learning
• Convolutional neural networks