Synopsis

A fully automatic, wrist cartilage segmentation method on magnetic resonance images is developed and validated. The method is based on convolutional neural networks (CNN). Cartilage segmentations obtained with the CNN showed a substantial agreement with manual segmentations for the full 3D wrist images and a good agreement for central coronal slices. The proposed method provided cartilage masks having a high concordance with manually obtained ones.

INTRODUCTION

A gold standard of the articular cartilage degradation assessment in patients with osteo- and rheumatoid arthritis is based on radiographic measurements of the joint space narrowing. This method is indirect and implies multiple measurements of distances between articular surfaces. Magnetic resonance imaging (MRI) allows the direct measurement of cartilage thickness. Manual and automatic tissue segmentation techniques can be used for cartilage extraction from 3D MR images to evaluate total cartilage volume of a studied joint^1,2. However, for such a complex joint as the wrist, the manual segmentation procedure becomes time consuming, and no automatic methods have been proposed for this purpose so far. A novel, deep learning-based method has shown a great potential for tissue segmentation in musculoskeletal MRI^3,4. In this study, we develop and validate an automatic procedure of wrist cartilage segmentation from 3D VIBE MR images based on a convolutional neural network (CNN).

METHODS

The study was conducted on 20 3D VIBE wrist images (acquired at 1.5T Magnetom Espree, Siemens) of 8 healthy volunteers and 3 patients suffered osteoarthritis. 3D images were converted into 1760 coronal slices, from which on 341 the cartilage was manually segmented by an experienced radiologist. The data were splitted into 4 datasets for the CNN training, development, test and validation (fig.1). The CNN input was a 28x28 pixels patch surrounding the pixel which had to be classified. The network architecture was constructed using Keras neural network library and had eight layers (2DConvolution, 2DMaxpooling, 4 2DConvolution, Flatten and Dense) and a two-class output. All the 3D data of subjects involved into the test phase were unseen by the CNN during the training phase. Dice coefficient (DSC) was calculated for the test data set. In order to validate the performance of CNN, a comparison with manual segmentation procedure results was held. A medial coronal slice of every 3D image (20 in total – validation dataset, also unseen by CNN) was segmented manually by 3 observers (twice by two of them, and once by one), and by the trained CNN. The manual segmentations made by the first observer in the first session were considered as a ground truth (GT). For the manual procedure, the intra- and interobserver concordance indices (CI) were calculated. The CNN-assessed segmentations were compared with GT through DSC and CI calculation.

RESULTS

The results of comparative analysis are summarized in Table 1. The trained CNN successfully segmented cartilage pixels as illustrated in Figure 2. One session of 20 slices manual segmentation of validation dataset made by a single observer took averagely about one hour, while the CNN processed the same dataset in 5 minutes.

DISCUSSION

The obtained here DSC value (0.73) is slightly lower than for the CNN-assisted knee cartilage segmentations (0.80-0.82)^3,4 . Taking into account a more complex structure of wrist joint and very limited training data, the obtained value can be considered as satisfactory. For the validation dataset the DSC reached 0.81, as it contained only medial coronal slices (fig.2 (a)) with high percentage of cartilage and higher contrast of cartilage pixels relatively to the surrounding tissues. In noisy images with low contrast of cartilage, the subcutaneous fat was occasionally considered by the CNN as cartilage (circled areas on fig. 2(b)). However, it still achieves higher concordance with the GT than the observers. Interestingly, the results obtained for both manual and CNN assisted segmentation showed similar CIs for healthy volunteers and patients. Images of patients contained pathological changes such as bone marrow edema (fig. 2(b)), cortex erosion (fig. 2(c)), ganglion cyst (fig. 2(d)). A vessel in a capitate bone of the second patient (fig. 2(d)) was not considered by CNN as cartilage. These results demonstrate a capability of the developed CNN to distinguish cartilage from other anatomical structures and possible joint abnormalities. The lack of input data for CNN training was a limitation of this study.

CONCLUSION:

Acknowledgements

References

1 Zink JV, Souteyrand P, Guis S et al. Standardized quantitative measurements of wrist cartilage in healthy humans using 3T magnetic resonance imaging. World Journal of Orthopedics 2015; 6(8): 641-648.

2 Peterfy CG, Van Dijke CF, Janzen DL et al. Quantification of articular cartilage in the knee with pulsed saturation transfer subtraction and fat-suppressed MR imaging: optimization and validation. Radiology 1994;192(2):485-91.

3 Prasoon A, Petersen K, Igel C et al. Deep feature learning for knee cartilage segmentation using a triplanar convolutional neural network. In: Med Image Comput Comput Assist Interv 2013;16(Pt 2):246-53.

4 Zhou Z, Zhao G, Kijowski R et al. Deep convolutional neural network for segmentation of knee joint anatomy. Magn Reson Med 2018;00:1–12.