Automatic estimation of knee effusion from limited MRI data

Abstract

AbstractKnee effusion is a common comorbidity in osteoarthritis. To quantify the amount of effusion, semi quantitative assessment scales have been developed that classify fluid levels on an integer scale from 0 to 3. In this work, we investigated the use of a neural network (NN) that used MRI Osteoarthritis Knee Scores effusion-synovitis (MOAKS-ES) values to distinguish physiologic fluid levels from higher fluid levels in MR images of the knee. We evaluate its effectiveness on low-resolution images to examine its potential in low-field, low-cost MRI. We created a dense NN (dNN) for detecting effusion, defined as a nonzero MOAKS-ES score, from MRI scans. Both the training and performance evaluation of the network were conducted using public radiological data from the Osteoarthritis Initiative (OAI). The model was trained using sagittal turbo-spin-echo (TSE) MR images from 1628 knees. The accuracy was compared to VGG16, a commonly used convolutional classification network. Robustness of the dNN was assessed by adding zero-mean Gaussian noise to the test images with a standard deviation of 5–30% of the maximum test data intensity. Also, inference was performed on a test data set of 163 knees, which includes a smaller test set of 36 knees that was also assessed by a musculoskeletal radiologist and the performance of the dNN and the radiologist compared. For the larger test data set, the dNN performed with an average accuracy of 62%. In addition, the network proved robust to noise, classifying the noisy images with minimal degradation to accuracy. When given MRI scans with 5% Gaussian noise, the network performed similarly, with an average accuracy of 61%. For the smaller 36-knee test data set, assessed both by the dNN and by a radiologist, the network performed better than the radiologist on average. Classifying knee effusion from low-resolution images with a similar accuracy as a human radiologist using neural networks is feasible, suggesting automatic assessment of images from low-cost, low-field scanners as a potentially useful assessment tool.