Sanders classification of calcaneal fractures in CT images with deep learning and differential data augmentation techniques.

BACKGROUND: Classification of the type of calcaneal fracture on CT images is essential in driving treatment. However, human-based classification can be challenging due to anatomical complexities and CT image constraints. The use of computer-aided classification system in standard practice is additionally hindered by the availability of training images. The aims of this study is to 1) propose a deep learning network combined with data augmentation technique to classify calcaneal fractures on CT images into the Sanders system, and 2) assess the efficiency of such approach with differential training methods. METHODS: In this study, the Principle component analysis (PCA) network was selected for the deep learning neural network architecture for its superior performance. CT calcaneal images were processed through PCA filters, binary hashing, and a block-wise histogram. The Augmentor pipeline including rotation, distortion, and flips was applied to generate artificial calcaneus fractured images. Two types of training approaches and five data sample sizes were investigated to evaluate the performance of the proposed system with and without data augmentation. RESULTS: Compared to the original performance, use of augmented images during training improved network performance accuracy by almost twofold in classifying Sanders fracture types for all dataset sizes. A fivefold increase in the number of augmented training images improved network classification accuracy by 35%. The proposed deep CNN model achieved 72% accuracy in classifying CT calcaneal images into the four Sanders categories when trained with sufficient augmented artificial images. CONCLUSION: The proposed deep-learning algorithm coupled with data augmentation provides a feasible and efficient approach to the use of computer-aided system in assisting physicians in evaluating calcaneal fracture types.

Deep learning and SURF for automated classification and detection of calcaneus fractures in CT images.

BACKGROUND AND OBJECTIVES: The calcaneus is the most fracture-prone tarsal bone and injuries to the surrounding tissue are some of the most difficult to treat. Currently there is a lack of consensus on treatment or interpretation of computed tomography (CT) images for calcaneus fractures. This study proposes a novel computer-assisted method for automated classification and detection of fracture locations in calcaneus CT images using a deep learning algorithm. METHODS: Two types of Convolutional Neural Network (CNN) architectures with different network depths, a Residual network (ResNet) and a Visual geometry group (VGG), were evaluated and compared for the classification performance of CT scans into fracture and non-fracture categories based on coronal, sagittal, and transverse views. The bone fracture detection algorithm incorporated fracture area matching using the speeded-up robust features (SURF) method, Canny edge detection, and contour tracing. RESULTS: Results showed that ResNet was comparable in accuracy (98%) to the VGG network for bone fracture classification but achieved better performance for involving a deeper neural network architecture. ResNet classification results were used as the input for detecting the location and type of bone fracture using SURF algorithm. CONCLUSIONS: Results from real patient fracture data sets demonstrate the feasibility using deep CNN and SURF for computer-aided classification and detection of the location of calcaneus fractures in CT images.

Improvement in the clinical practicability of roentgen stereophotogrammetric analysis (RSA): free from the use of the dual X-ray equipment.

After total knee replacement, the monitoring of the prosthetic performance is often done by roentgenographic examination. However, the two-dimensional (2D) roentgen images only provide information about the projection onto the anteroposterior (AP) and mediolateral (ML) planes. Historically, the model-based roentgen stereophotogrammetric analysis (RSA) technique has been developed to predict the spatial relationship between prostheses by iteratively comparing the projective data for the prosthetic models and the roentgen images. During examination, the prosthetic poses should be stationary. This should be ensured, either by the use of dual synchronized X-ray equipment or by the use of a specific posture. In practice, these methods are uncommon or technically inconvenient during follow-up examination. This study aims to develop a rotation platform to improve the clinical applicability of the model-based RSA technique. The rotation platform allows the patient to assume a weight-bearing posture, while being steadily rotated so that both AP and ML knee images can be obtained. This study uses X-ray equipment with a single source and flat panel detectors (FPDs). Four tests are conducted to evaluate the quality of the FPD images, steadiness of the rotation platform, and accuracy of the RSA results. The results show that the distortion-induced error of the FPD image is quite minor, and the prosthetic size can be cautiously calibrated by means of the scale ball(s). The rotation platform should be placed closer to the FPD and orthogonal to the projection axis of the X-ray source. Image overlap of the prostheses can be avoided by adjusting both X-ray source and knee posture. The device-induced problems associated with the rotation platform include the steadiness of the platform operation and the balance of the rotated subject. Sawbone tests demonstrate that the outline error, due to the platform, is of the order of the image resolution (= 0.145 mm). In conclusion, the rotation platform with steady rotation, a knee support, and a handle can serve as an alternative method to take prosthetic images, without the loss in accuracy associated with the RSA method.

Improved execution efficiency of model-based roentgen stereophotogrammetric analysis: simplification and segmentation of model meshes.

Recently, the model-based roentgen stereophotogrammetric analysis (RSA) method has been developed as an in vivo tool to estimate static pose and dynamic motion of the instrumented prostheses. The two essential inputs for the RSA method are prosthetic models and roentgen images. During RSA calculation, the implants are often reversely scanned and input in the form of meshes to estimate the outline error between prosthetic projection and roentgen images. However, the execution efficiency of the RSA iterative calculation may limit its clinical practicability, and one reason for inefficiency may be very large number of meshes in the model. This study uses two methods of mesh manipulation to improve the execution efficiency of RSA calculation. The first is to simplify the model meshes and the other is to segment and delete the meshes of insignificant regions. An index (i.e. critical percentage) of an optimal element number is defined as the trade-off between execution efficiency and result accuracy. The predicted results are numerically validated by total knee prosthetic system. The outcome shows that the optimal strategy of the mesh manipulation is simplification and followed by segmentation. On average, the element number can even be reduced to 1% of the original models. After the mesh manipulation, the execution efficiency can be increased about 75% without compromising the accuracy of the predicted RSA results (the increment of rotation and translation error: 0.06° and 0.02 mm). In conclusion, prosthetic models should be manipulated by simplification and segmentation methods prior to the RSA calculation to increase the execution efficiency and then to improve clinical applicability of the RSA method.

Automatic model-based roentgen stereophotogrammetric analysis (RSA) of total knee prostheses.

Conventional radiography is insensitive for early and accurate estimation of the mal-alignment and wear of knee prostheses. The two-staged (rough and fine) registration of the model-based RSA technique has recently been developed to in vivo estimate the prosthetic pose (i.e, location and orientation). In the literature, rough registration often uses template match or manual adjustment of the roentgen images. Additionally, possible error induced by the nonorthogonality of taking two roentgen images neither examined nor calibrated prior to fine registration. This study developed two RSA methods for automate the estimation of the prosthetic pose and decrease the nonorthogonality-induced error. The predicted results were validated by both simulative and experimental tests and compared with reported findings in the literature. The outcome revealed that the feature-recognized method automates pose estimation and significantly increases the execution efficiency up to about 50 times in comparison with the literature counterparts. Although the nonorthogonal images resulted in undesirable errors, the outline-optimized method can effectively compensate for the induced errors prior to fine registration. The superiority in automation, efficiency, and accuracy demonstrated the clinical practicability of the two proposed methods especially for the numerous fluoroscopic images of dynamic motion.

A new registration method for three-dimensional knee nearthrosis model using two X-ray images.

The purpose of this study is to develop a method to analyse the pose of the knee nearthrosis mounted and to automate the registration procedure for easy use in clinical applications. The proposed registration method is essentially a model-based method, in which the CAD model is acquired by reverse engineering. The CAD model is converted into a two-dimensional (2D) image by a rendering technique, and the compatibility of the X-ray image and the image of the CAD model is investigated. To avoid the optimisation of six unknown parameters with respect to the relative pose between the condyle and tibial models, a 2D coordinate system is set on each component of the X-ray images. A 3D coordinate system is also set on each of the two nearthrosis components. With such a setup, there is only one unknown rotational angle on each component, which is determined by an optimum algorithm in accordance with the contour error between the X-ray image and the image of the CAD model. Extensive computer simulation and in vitro experiments using real X-ray images have been implemented to investigate the feasibility of the proposed registration method.