Introduction

Quantitative compositional MRI biomarkers, such as T₂ relaxation time mapping, play a central role in Osteoarthritis(OA) research, probing the biochemical composition of the articular cartilage¹. Current standard is to measure averages T₂ values in manually defined cartilage compartments², however there is growing interest in exploring techniques for the automatic analysis of spatial distribution and extraction of relaxometry patterns able to characterize subjects and predict disease progression³. Recently, deep learning has dramatically improved some of the most challenging artificial intelligence and medical informatics tasks including drug discovery and genomics⁴. Deep learning computational models are composed of multiple processing layers and learn representations of data with multiple levels of abstraction using the tendency that many natural patterns are compositional hierarchies. In this study we coupled T₂ relaxation time compositional imaging and deep convolutional neural networks with the aim to automatically extract relaxometry patterns and data-driven features able to: (i) distinguish OA subjects from controls (ii) predict incidence of OA 4 years before any radiographic signs.

Method

1,348 subjects from the Osteoarthritis Initiative (OAI) public dataset were considered in this study. 3 groups were defined by the evaluation of Kellgren and Lawrence (KL) scores at baseline (V₀) and 4 year (V₄) time points: the Control Cohort V₀ KL 0-1; V₄ KL 0-1 (N=612(48.29%), age 58.39 years, BMI 27.94 Kg/m², 310 female), the OA Cohort V₀ KL ≥2; V₄ KL ≥2 (N=651(45.4%), age 63.48 years BMI 30.34 Kg/m², 438 female) and the Progression Cohort V₀ KL 0-1; V₄ KL ≥2 (N=85(6.3%), age 59.28 years, BMI 29.84 Kg/m², 54 female). Sagittal 2-D multi-echo spin-echo images were used for the quantification of the T₂ relaxation time. All the T₂ maps were morphed to a common standard space, performing a fully automatic atlas based cartilage segmentation previously described⁵ (Figure 1A). The morphing of all the T₂ relaxation time maps allowed the description of each map with a matched vector of 11,980 components building a multidimensional signature that capture all the cartilage compositional information included in each T₂ map (Figure 1B). Each vector was then locally normalized considering matched controls groups, obtaining a description invariant to demographics covariates and global spatial pattern that captures just anomalies (Figure 1C). Unsupervised data exploration based on Topological Data Analysis (TDA)⁶ was performed for the multidimensional visualization of the dataset with the aim of evaluating the voxel based T₂ relaxation time N-Dimensional description as a valid osteoarthritis feature space (Figure 2). All the normalized maps were then converted in 2D images (Figure 3) by mapping the T₂ values using a cartilage flattening technique aimed to preserved the structural information included in the maps (cartilage layers, directional heterogeneity). These 2D maps were used for the training of a deep convolutional neural network described in Figure 4. Due to the small number of subjects in the progression group, simplified feature maps including either medial or lateral information were explored to accomplish the second aim of this study. 25% of the subjects in each class were randomly selected as validation set and not used in the training phase. Principal Component Analysis (PCA)-based technique was used for data augmentation⁷ of the remaining 75% with the aim of synthetizing feasible maps and to increase the sample size to 13,480 samples.

Results

The automatic method adopted in this study produced T₂ compartmental averages strongly correlated to the ones obtained by manual segmentation processing (R =0.84, Figure 5). The binary classification between OA and controls subjects showed a classification accuracy of 95.2% when applied in the validation test. As a comparison, when the 4 compartmental averages obtained from manual process (LF, MF, LT, MT) were considered to train a Support Vector Machine (SVM) the accuracy was just 59.18%, showing the power of applying a deep learning model and extracting data driven features from the overall map. The progression prediction showed the best performances when applied on the medial femur compartment alone with a percentage of correct classification in the validation set equal to 80.7%.

Discuss and Conclusion

In this study we provide a proof of concept that the coupling of cutting edge technologies in quantitative compositional MRI and deep learning fields can successfully discover latent feature representations, non-linear aggregation among elementary features able to accurately characterize disease status and predict progression. The data-driven extraction of features from T₂ maps can exploit the real potential of this quantitative MRI technique, to date still hampered by tedious and time-consuming manual image post-processing pipelines; and deeply underused due to the handcrafting of too simplistic image representations, such as relaxation averages or first order texture information.

Acknowledgements

P50 AR060752 (SM), R01AR046905 (SM), H. Neilsen Foundation (JH and ARF), R01NS067092 (ARF), R01NS088475 (ARF), Wings for Life (ARF),

References

1. Li X, Majumdar S. Quantitative MRI of articular cartilage and its clinical applications. Journal of magnetic resonance imaging 2013;38:991-1008. 2. Joseph GB, McCulloch CE, Nevitt MC, et al. A reference database of cartilage 3 T MRI T2 values in knees without diagnostic evidence of cartilage degeneration: data from the osteoarthritis initiative. Osteoarthritis and Cartilage 2015;23:897-905. 3. Pedoia V, Russell C, Randolph A, Li X, Majumdar S. Principal component analysis-T1r Voxel based relaxometry of the articular cartilage: a comparison of biochemical patterns in osteoarthritis and anterior cruciate ligament subjects. Quant Imaging Med Surg. 2016 Dec; 6(6):623-633. 4. LeCun Y, Bengio Y, Hinton G. Deep learning. Nature 2015;521:436-44. 5 Pedoia V, Li X, Su F, Calixto N, Majumdar S. Fully automatic analysis of the knee articular cartilage T1rho relaxation time using voxel-based relaxometry. Journal of magnetic resonance imaging : JMRI 2016;43:970-80 6. Lum PY, Singh G, Lehman A, et al. Extracting insights from the shape of complex data using topology. Nature Scientific reports 2013;3:1236. 7. A Krizhevsky, I Sutskever, GE Hinton ImageNet Classification with Deep Convolutional Neural Networks Advances in neural information processing systems, 1097-1105