Manual segmentation of articular cartilage and menisci from magnetic resonance data is time-consuming and can be challenging. Recently, deep learning has shown promising results in medical image segmentation. The aim of this study was to develop a method for automatic segmentation of articular cartilage and menisci that performs more accurately and efficiently than the previously published methods. On OAI/iMorphics dataset the method achieves Dice score of 90.7±1.9 for femoral cartilage, 89.7±2.8 for tibial, 87.1±4.7 for patellar, and 86.3±3.4 for menisci. The presented results could facilitate the osteoarthritis research and enhance clinical practice. Source codes and pretrained models will be open-sourced.
Osteoarthritis (OA) is the most common musculoskeletal disease in the world. Magnetic resonance imaging (MRI) has been actively used to assess the condition of the affected tissues and to study the onset and progression of OA. 3D double-echo steady-state (3D-DESS) MRI sequence provides a clear contrast between cartilage, bone, meniscus and synovial tissues, which is generally sufficient for assessment of tissue lesions by a radiologist. Automatic segmentation of articular cartilage and menisci is of high relevance for OA diagnostics and basic research, but it still remains a challenging task. Once solved, it can lead to a significant reduction of time required for the automatic analysis of the MRI data.
During the past years, deep learning (DL) has become a gold standard in medical image segmentation. However, despite multiple attempts, its value in the task of segmentation of cartilage and meniscal tissue has not been fully clarified. In this study, we introduce a fully-automatic method to segment the cartilage tissue and menisci from 3D-DESS MRI data, demonstrating the improvement over the previously published methods.
We utilized a subset of Osteoarthritis Initiative (OAI) dataset[1] and the corresponding annotations produced by iMorphics[2]. The data consisted of sagittal 3D-DESS MRI scans from 88 subjects (2 scans per subject: at baseline and at 12 months), acquired with Siemens 3.0T MRI scanners. Each scan contained 160 slices of 384x384 pixels (slice thickness = 0.7mm, field of view = 14x14 cm2). From these data, the following tissues were segmented: femoral cartilage, tibial cartilage (lateral and medial as one class), patellar cartilage, and menisci (lateral and medial as one class).
We randomly split the whole dataset subject-wise into train and test subsets maintaining the similar distribution of the Kellgren-Lawrence (KL) gradings (derived from the corresponding radiographs) between the subsets. Subsequently, we used 5-fold stratified cross-validation splits of the train data to develop the segmentation models. The test set was used solely for the final stage of our analysis.
Our segmentation approach is based on U-Net[3] with several modifications (24 channels in the first convolutional block output, doubled at each level, 6 levels total), performing cartilage tissue segmentation slice-wise. We trained the same model for each cross-validation fold for 50 epochs using multi-class cross-entropy loss function and Adam optimizer. The models were trained with a starting learning rate of 0.001, reduced by the factor of 10 at 30th epoch, weight decay of 5e-5, and batch size of 48. To improve the generalization of our solution, during the training we used a set of data augmentations: histogram percentile clipping (lower 5% and upper 1% of sample intensity), horizontal flips, gamma correction, artificial worsening (downscaling followed by upscaling), and bilateral filtering. For training the models, we used PyTorch DL framework and 2xNvidia GTX 1080 graphical processing units.
In the testing phase, we produced the predictions using all the 5 trained models and averaged their outputs. To assess the results, we used volumetric Dice score coefficient (DSC). In order to gain more insights regarding the performance of our approach, and assuming the data being registered, we calculated the statistics of DSC for each spatial slice across the test subset scans.