Fig 5 - uploaded by Tomasz Gawron
Content may be subject to copyright.
Example results from validation set. Even when keypoint positions are not precisely estimated, the arc they are representing corresponds to the actual feature of the bone.
Source publication
A path along which the human knee joint moves can be estimated from real-time moving images or a sequence of static images. In case of many algorithms solving this problem, it is essential to locate the characteristic points (i.e., key-points) on each image and find the correspondence between them in the image sequence. In this paper we present an...
Contexts in source publication
Context 1
... example predicted positions of key-points, together with corresponding manually assigned points, are presented in Fig. 5 and Fig. 6. The images show the input of the CNN, i.e. the windows of size 178px×178px. Blue markers denote key-points obtained on the basis of manually assigned points, red markers represent key-points estimated by CNN. The key-points positions are estimated with an acceptable precision even when the image is excessively cropped. Three ...
Context 2
... example predicted positions of key-points, together with corresponding manually assigned points, are presented in Fig. 5 and Fig. 6. The images show the input of the CNN, i.e. the windows of size 178px×178px. Blue markers denote key-points obtained on the basis of manually assigned points, red markers represent key-points estimated by CNN. The key-points positions are estimated with an acceptable precision even when the image is excessively cropped. Three key-points can be simultaneously estimated with decent precision by a single network. Since the femur tracking is based on an arc representing PS, not the key-points itself, when key-point estimate is slightly inaccurate, the resulting arc fits the ground truth with satisfactory ...
Citations
... Thus, keypoints reflect the configuration of PS on the source image. As a first stage, due to the small dataset size, the original data were augmented with typical image transformations (rotation, translation, scale, reflection, contrast change [26]). Second, image frames were cropped to size 178 × 178 px. ...
... Nevertheless, we are convinced that proper regularization will assure an appropriate learning process. Following the work presented in [26], we have applied: ...
... The network architecture is presented in Figure 8. The optimal CNN architecture [26] consists of 15 layers, 10 of which are convolutional. The size of the last layer represents the number of network outputs, i.e., the coordinates of keypoints k 1 , k 2 , k 3 . ...
In this paper, we present a fully automatic solution for denoting bone configuration on two-dimensional images. A dataset of 300 X-ray images of children’s knee joints was collected. The strict experimental protocol established in this study increased the difficulty of post-processing. Therefore, we tackled the problem of obtaining reliable information from medical image data of insufficient quality. We proposed a set of features that unambiguously denoted configuration of the bone on the image, namely the femur. It was crucial to define the features that were independent of age, since age variability of subjects was high. Subsequently, we defined image keypoints directly corresponding to those features. Their positions were used to determine the coordinate system denoting femur configuration. A complex keypoint detector was proposed, composed of two different estimator architectures: gradient-based and based on the convolutional neural network. The positions of the keypoints were used to determine the configuration of the femur on each image frame. The overall performance of both estimators working in parallel was evaluated using X-ray images from the publicly available LERA dataset.
