A Deep-Learning Approach to Detect Fiducials in Planar X-Ray Images for Undistortion of Conventional C-Arm Images
In some applications such as 2D-3D registration, undistorted images are required to achieve optimal results. These types of images can be obtained from a distortion-free C-arm (flat-panel detector) or by undistorting the images given from a conventional C-arm (analogue image intensifier.) Undistorti...
Guardado en:
| Autores principales: | , , , |
|---|---|
| Formato: | article |
| Lenguaje: | EN |
| Publicado: |
De Gruyter
2020
|
| Materias: | |
| Acceso en línea: | https://doaj.org/article/4bb757a8f10f43879e031055bfa4a21b |
| Etiquetas: |
Agregar Etiqueta
Sin Etiquetas, Sea el primero en etiquetar este registro!
|
| Sumario: | In some applications such as 2D-3D registration, undistorted images are required to achieve optimal results. These types of images can be obtained from a distortion-free C-arm (flat-panel detector) or by undistorting the images given from a conventional C-arm (analogue image intensifier.) Undistorting images require a plate with fiducials connected to the C-arm detector. Detecting fiducials is affected by differences in the image contrast due to elements in the background. Therefore, the results vary from image to image and could require manual tuning of parameters. We propose a deep-learning approach for detecting undistortion-platefiducials in X-ray images to overcome the drawbacks previously stated. With an undistortion plate, we took 1120 XRays using a C-arm in different poses. Every X-ray is afterward cut into 60 sub-images. We used these sub-images for training a convolutional neural network (CNN). Comparing the CNN and a traditional image processing method based on the Hough Circle algorithm, we found that the detected fiducials using the traditional method give a similar fiducial positioning error. Nevertheless, the fiducial detection rate goes from 89.7% using the traditional method to 100% with the developed CNN. The results show that the detection rate and precision of our deep-learning approach guarantee the undistortion of conventional C-Arm images. |
|---|