Synopsis

Keywords: Machine Learning/Artificial Intelligence, Joints

We evaluated three neural network architectures for the automatic identification of the center of the femur head on 3D water-only Dixon MRI. We trained using a mixture of real and augmented data. The mean error of the best-performing network was three-time lower compared to a manual annotation and on the order of 1 voxel. We combined the network to create the first fully automated pipeline to assess the hip range of motion from 3D MR.

Introduction

Femoroacetabular impingement (FAI) is a pathological condition of the hip joint characterized by an abnormal shape of the bones that results in cartilage erosion from contact between the acetabular rim and femoral head-neck junction during hip motion[1-3]. Computed Tomography (CT) has been used to create patient-specific simulation models to determine the contact points during hip motion and assess the joint range of motion (ROM)[3]. However, CT results in potentially harmful ionizing radiations in the pelvis area[4]. It has been recently shown that CT could be replaced by magnetic resonance imaging (MRI)[3], by means of a 3D Dixon-based pulse sequence that enables the automatic segmentation of bones[4,5]. All previous approaches need user intervention to identify the center of the femur head (CFH), which is commonly used as the center of rotation of the joint. This study focused on the development of a neural network to automatically find the CFH, with the overall goal of creating a fully automated pipeline for the evaluation of hip ROM (Figure 1).

Material and Methods

For this study, we used data from 40 hip joints (left and right) of 17 patients (13F/4M,37.1±5.7years), who underwent an MRI examination before arthroscopic surgery. Three patients had a follow-up MRI after one year. The study was approved by the IRB. The exam included a 3D Dixon MRI of the pelvis [8], acquired with a FOV = 320x320 mm2,320x320 matrix size,1 mm slice thickness without gap, centered on the patient's pelvis. The femur and acetabulum were automatically segmented using a pre-trained U-Net network[5,6] on the water-only Dixon images and the results were checked by an expert radiologist. After that, the CFH of each joint was estimated as the center of a sphere fitted in a semi-automatic way to a 3D mesh of the femoral head using an ad-hoc script(Figure 2). The obtained CFH was assumed as the ground truth for training. For comparison, the CFH also annotated manually the MRI data by clicking a point on an axial section through the femoral head. Images, segmented regions of interest (ROIs), and the CFH coordinates were used to train three variations of a LeNET architecture (Figure 3) with an input size of 80x80x80 voxels. The first architecture (LeNET) takes as input the images and the ROIs and gives as output an array with three values representing the CFH coordinates x, y, and z. The second architecture (LeNETv1) takes as input the images while the ROIs are concatenated after the first convolution so that the segmentation becomes one of the feature maps of the network for the second convolutional layer and the same output of LeNET. The third one (LeNET3) has the same input as LeNET, but gives three outputs, each with a scalar value representing the three coordinates (x,y,z) of the CFH. The 40 hip datasets were randomly subdivided into 30 (15 patients) training sets and 10 validation sets. Images and ROIs were normalized by their maximum and resized to a field of view of 80x80x80 mm3 and 80x80x80 voxels to fit the network input. All three networks were trained using 320 Epochs, batch size = 16, Adam optimizer (lr 0.001), and the mean of the Euclidean distance of CFH as a loss. Every batch was composed of 4 real and 12 augmented datasets. The latter was created on the fly by transforming the MRI, ROIs, and CFH coordinates using 6 rototranslations randomly selected from a uniform distribution ranging between ±5 degrees for the two first Euler’s angles (left/right axes and anterior/posterior), ±15 for the (inferior/superior) and ±5 mm for translations.

Results

Figure 4 shows that LeNET and LeNETv1 had a median error of fewer than 1 mm2 while the performance of LeNET3 was significantly worse (p<0.01). In particular, the mean error for LeNET over the 10 validation data was 0.95 mm(±0.43 mm), for LeNETv1 was 1.11 mm(±0.66 mm), and for LeNET3 was 40.01 mm (±0.01 mm).

Discussion

Among the proposed networks, LeNET had the best performance. Less than 1 mm is comparable to the spatial resolution of the images and considerably better than the error obtained by manually selecting the CFH on MRI slices, which was more than 2.9 mm(±1.44). In the example in Figure 5, the femur and acetabulum are automatically segmented from the water-only Dixon images of a representative patient[5], a 3D mesh of the bones is generated, LeNET identifies the CFH, which is used as the center of rotation to simulate the hip ROM buy detecting collisions between the head-neck junction of the femur and the rim of the acetabulum. The final ROM is estimated statistically by repeating the analysis 100 times while uniformly varying the CFH coordinates by 1 mm in space, to reliably overcome the average error of the LeNET network.

Conclusion

Despite the limited number of data available, this work demonstrates that it is possible to automatically estimate the CFH to create an automatic pipeline for the 3D simulation of the hip ROM using MRI data. In future work, we will increase the number of training data to further improve the accuracy of the CFH estimation and indirectly the reliability of the ROM assessment.

Acknowledgements

his work was supported by NIH R01 AR070297 and performed under the rubric of the Center for Advanced Imaging Innovation and Research (CAI2R, www.cai2r.net), an NIBIB National Center for Biomedical Imaging and Bioengineering (NIH P41 EB017183).