Synopsis

Automated knee cartilage segmentation can potentially improve the clinical utility of the MRI assessment of knee osteoarthritis due to the convoluted structure of the knee cartilage in 3D. Recently deep convolutional neural network (CNN) have shown better performance for knee cartilage segmentation. Unlike other segmentation algorithms deep-CNN techniques learn the model parameters from the data itself. Therefore, this abstract proposed that deep 3D-CNN techniques can be used to determine the optimal MRI sequence for knee cartilage segmentation and demonstrated that 3D-DESS MRI have statistically better segmentation performance as compared to 3D-T₁-FLASH MRI.

Introduction

Method

Patient data was obtained from Osteoarthritis initiative (OAI)⁹. 88 patients with one baseline and one 12 month follow-up MRI scan was used with labels (femoral cartilage, patellar cartilage, left and right tibial cartilage and left and right menisci) annotated over sagittal 3D-DESS MRI (160 slices, 0.7mm slice thickness, water excitation, FOV of 140mmX140mm, in-plane resolution 0.365mmX0.456mm, TR/TE=16.3ms/4.7ms, flip angle = 25^o). Labels were transferred from Sagittal 3D-DESS MRI to Coronal 3D-T₁-FLASH MRI (80 slices, 1.5mm slice thickness, water excitation, FOV of 160mmX160mm, in-plane resolution 0.3125mmX0.3125mm, TR/TE=20ms/7.57ms, flip angle = 12^o).

Deep 3D-CNN network called µ -Net⁸ (Figure 1) was used to independently train and test on sagittal-DESS and Coronal T₁-FLASH MRI. µ-Net is motivated by 3D U-Net¹⁰ and 3D-V-Net¹¹. In addition µ-Net has, Short-skip and long-skip connections to carry out the flow of information within the network, Element wise addition instead of concatenation operation is used in skip connection to reduce memory footprint on the GPU, Input is fed at multiple-resolution in the analysis path to minimize the information lost during down-sampling and an Auxiliary loss layer is added at each step in the synthesis part of the network, to provide a form of deep supervision. Various error metrics (Dice Score (DSC), Volume Overlap Error (VOE), Volume Difference (VD) and Hausdorff Distance of surfaces (ED)) were adopted for comparison of the predicted segmentation results (S) with the reference ground truth segmentation (R),

$$DSC =\frac{2\left |S\cap R \right |}{\left |S\right |+\left |R\right |}$$,

$$VOE =\frac{1-\left |S\cap R\right |}{\left |S\cup R \right |}$$,

$$VD =\frac{\left |S\right |-\left |R\right |}{\left |R\right |}$$

for femoral cartilage, patellar cartilage, left and right tibial cartilage and left and right menisci. Five-fold cross validation with 140 training and 36 test datasets was used to compute the 95% confidence interval (CI) of the measured error metrics. No overlap between the 95% CI was used as a measure for statistically significant difference with p-value of 0.05 between segmentation obtained from DESS and T₁-FLASH 3D-MRI.

Results and Discussion

Conclusion

Acknowledgements

References

[1] Cross, M., Smith, E., Hoy, D., et. al., The global burden of hip and knee osteoarthritis: estimates from the global burden of disease 2010 study, Annals of the rheumatic diseases, 2014.

[2] Gray, M.L., Eckstein, F., Peterfy, C et. al., Toward imaging biomarkers for osteoarthritis, Clinical orthopaedics and related research. 2004; S175-S181.

[3] Eckstein, F., Burstein, D. and Link, T.M., Quantitative MRI of cartilage and bone: degenerative changes in osteoarthritis. NMR in Biomedicine. 2006; 19(7): 822-854.

[4] Eckstein, F., Hudelmaier, M., Wirth, W. et. al. Double echo steady state magnetic resonance imaging of knee articular cartilage at 3 Tesla: a pilot study for the Osteoarthritis Initiative. Annals of the rheumatic diseases. 2006; 65(4): 433-441.

[5] Shim, H., Chang, S., Tao, C. et. al., Knee Cartilage: Efficient and Reproducible Segmentation on High-Spatial-Resolution MR Images with the Semiautomated Graph-Cut Algorithm Method, Radiology : 2009, 251(2); 548-556.

[6] Heimann, T., Morrison, B.J., Styner, M.A. et al., Segmentation of knee images: A grand challenge. Proc. MICCAI Workshop on Medical Image Analysis for the Clinic. 2010 : 207-214.

[7] Prasoon, A., Petersen, K., Igel, C. et al., Deep feature learning for knee cartilage segmentation using a triplanar convolutional neural network. MICCAI, 2013 : 246-253

[8] Raj A., Ajani B., Krishnan K., et. al., 3D Deep Learning Based Cartilage Segmentation for Visualization and Analysis of Quantitative MRI for Osteoarthritis. Proc. of 10th ISMRM Workshop on Osteoarthritis Imaging, 2017.

[9] Peterfy, C.G., Schneider, E. and Nevitt, M., The osteoarthritis initiative: report on the design rationale for the magnetic resonance imaging protocol for the knee. Osteoarthritis and cartilage, 2008 (16.12): 1433-1441.

[10] Çiçek, Ö., Abdulkadir, A., Lienkamp, S.S. et al., 3D U-Net: learning dense volumetric segmentation from sparse annotation. International Conference on MICCAI, 2016:424-432.

[11] Milletari, F., Navab, N. and Ahmadi, S.A., V-net: Fully convolutional neural networks for volumetric medical image segmentation, 3D Vision (3DV), 2016 Fourth International Conference on. IEEE, 2016 : 565-571.