A discriminative shape-texture convolutional neural network for early diagnosis of knee osteoarthritis from X-ray images

Knee Osteoarthritis (OA) is one of the most common causes of physical disability worldwide associated with a significant personal and socioeconomic burden. Deep Learning approaches based on Convolutional Neural Networks (CNNs) achieved remarkable improvements in knee OA detection. Despite this success, the problem of early knee OA diagnosis from plain radiographs remains a challenging task. This is due to the high similarity between the X-ray images of OA and non-OA subjects and the disappearance of texture information regarding bone microarchitecture changes in the top layers during the learning process of the CNN models. To address these issues, we propose a Discriminative Shape-Texture Convolutional Neural Network (DST-CNN), which automatically diagnoses early knee OA from X-ray images. The proposed model incorporates a discriminative loss to improve class separability and deal with high inter-class similarities. In addition, a new Gram Matrix Descriptor (GMD) block is embedded in the CNN architecture to compute texture features from several intermediate layers and combine them with the shape features in the top layers. We show that merging texture features with deep ones leads to better prediction of the early stages of OA. Comprehensive experimental results on two large public databases, Osteoarthritis Initiative (OAI) and Multicenter Osteoarthritis Study (MOST) demonstrate the potential of the proposed network. Ablation studies and visualizations are provided for a detailed understanding of our proposed approach.

Discriminative Regularized Auto-Encoder for Early Detection of Knee OsteoArthritis: Data from the Osteoarthritis Initiative

OsteoArthritis (OA) is the most common disorder of the musculoskeletal system and the major cause of reduced mobility among seniors. The visual evaluation of OA still suffers from subjectivity. Recently, Computer-Aided Diagnosis (CAD) systems based on learning methods showed potential for improving knee OA diagnostic accuracy. However, learning discriminative properties can be a challenging task, particularly when dealing with complex data such as X-ray images, typically used for knee OA diagnosis. In this paper, we introduce a Discriminative Regularized Auto Encoder (DRAE) that allows to learn both relevant and discriminative properties that improve the classification performance. More specifically, a penalty term, called discriminative loss is combined with the standard Auto-Encoder training criterion. This additional term aims to force the learned representation to contain discriminative information. Our experimental results on data from the public multicenter OsteoArthritis Initiative (OAI) show that the developed method presents potential results for early knee OA detection.

Anisotropic Discrete Dual-Tree Wavelet Transform for Improved Classification of Trabecular Bone

This paper deals with a new anisotropic discrete dual-tree wavelet transform (ADDTWT) to characterize the anisotropy of bone texture. More specifically, we propose to extend the conventional discrete dual-tree wavelet transform (DDTWT) by using the anisotropic basis functions associated with the hyperbolic wavelet transform instead of isotropic spectrum supports. A texture classification framework is adopted to assess the performance of the proposed transform. The generalized Gaussian distribution is used to model the distribution of the sub-band coefficients. The estimated vector of parameters for each image is then used as input for the support vector machine classifier. Experiments were conducted on synthesized anisotropic fractional Brownian motion fields and on a real database composed of osteoporotic patients and control cases. Results show that the ADDTWT outperforms most of the competing anisotropic transforms with an area under curve rate of 93%.

3D shape-dependent thinning method for trabecular bone characterization

PURPOSE

Curve and surface thinning are widely-used skeletonization techniques for modeling objects in three dimensions. In the case of trabecular bone analysis, however, neither curve nor surface thinning is really efficient since the internal geometry of the object is usually composed of both rod and plate shapes. The purpose of this paper is to propose an original method called hybrid skeleton which better matches the geometry of the data compared to curve and surface skeletons. In the hybrid skeleton algorithm, 1D curves represent rod-shaped zones whereas 2D surfaces represent plate-shaped elements.

METHODS

The proposed hybrid skeleton algorithm is based on a combination of three methods. (1) A new variant of the method proposed by Bonnassie et al. for the classification of voxels as belonging to plate-like or rod-like structures, where the medial axis (MA) algorithm is replaced by a fast and connected skeletonization algorithm. In addition, the reversibility of the MA algorithm is replaced by an isotropic region-growth method to spread the rod and plate labels back to the original object. (2) A well chosen surface thinning method applied on the plate voxels set. (3) A well chosen curve skeleton thinning method applied on the rod voxels set. The efficiency and the robustness of the proposed algorithm were evaluated using synthesis test vectors. A clinical study was led on micro-CT (computed tomography) images of two different populations of osteoarthritic and osteoporotic trabecular bone samples. The morphological and topological characteristics of the two populations were evaluated using the proposed hybrid skeleton as well as the classification algorithm.

RESULTS

When evaluated on test vectors and compared to Bonnassie's algorithm, the proposed classification algorithm gives a slightly better rate of classification. The hybrid skeleton preserves the shape information of the processed objects. Interesting morphological and topological features as well as volumetric ones were extracted from the skeleton and from the classified volumes, respectively. The extracted features enable the two populations of osteoarthritic and osteoporotic trabecular bone samples to be distinguished.

CONCLUSIONS

Compared to curve-based or surface-based skeletons, the hybrid skeleton better matches the geometry of the data. Each rod is represented by a one-voxel-thick arc and each plate is represented by a one-voxel-thick surface. The hybrid skeleton as well as the proposed classification algorithm introduce relevant parameters linked to the presence of plates in the trabecular bone data, showing that rods and plates contain independent information about trabeculae. The hybrid skeleton offers a new opportunity for precise studies of porous media such as trabecular bone.

Genetic algorithm and image processing for osteoporosis diagnosis

Osteoporosis is considered as a major public health threat. It is characterized by a decrease in the density of bone, decreasing its strength and leading to an increased risk of fracture. In this work, the morphological, topological and mechanical characteristics of 2 populations of arthritic and osteoporotic trabecular bone samples are evaluated using artificial intelligence and recently developed skeletonization algorithms. Results show that genetic algorithms associated with image processing tools can precisely separate the 2 populations.

Estimation of the 3D self-similarity parameter of trabecular bone from its 2D projection

It has been shown that the analysis of two dimensional (2D) bone X-ray images based on the fractional Brownian motion (fBm) model is a good indicator for quantifying alterations in the three dimensional (3D) bone micro-architecture. However, this 2D measurement is not a direct assessment of the 3D bone properties. In this paper, we first show that S(3D), the self-similarity parameter of 3D fBm, is linked to S(2D), that of its 2D projection, by S(3D)=S(2D)-0.5. In the light of this theoretical result, we have experimentally examined whether this relation holds for trabecular bone. Twenty one specimens of trabecular bone were derived from frozen human femoral heads. They were digitized using a high resolution mu-CT. Their projections were simulated numerically by summing the data in the three orthogonal directions and both 3D and 2D self-similarity parameters were measured. Results show that the self-similarity of the 3D bone volumes and that of their projections are linked by the previous equation. This demonstrates that a simple projection provides 3D information about the bone structure. This information can be a valuable adjunct to the bone mineral density for the early diagnosis of osteoporosis.

Fractal analysis of bone X-ray tomographic microscopy projections

Fractal analysis of bone X-ray images has received much interest recently for the diagnosis of bone disease. In this paper, we propose a fractal analysis of bone X-ray tomographic microscopy (XTM) projections. The aim of the study is to establish whether or not there is a correlation between three-dimensional (3-D) trabecular changes and two-dimensional (2-D) fractal descriptors. Using a highly collimated beam, 3-D bone X-ray tomographic images were obtained. Trabecular bone loss was simulated using a mathematical morphology method. Then, 2-D projections were generated in each of the three orthogonal directions. Finally, the model of fractional Brownian motion (fBm) was used on bone XTM 2-D projections to characterize changes in bone structure that occur during disease, such a simulation of bone loss. Results indicate that fBm is a robust texture model allowing quantification of simulations of trabecular bone changes.