Slice Interpolation for Medical Image based on Spatial Geometry Polynomial Fitting

Introduction: Medical imaging technologies are essential to disease diagnosis and surgery planning, such as computational tomography (CT), which expresses the acquired tomographic medical image data as a set of slice sequences. In order to decrease the radiation amount received by the patients, it is a common practice to reduce the sampling rate and improve the scanning speed, which resulting in the loss of some valuable temporal information and remarkably large slice interval. Therefore, most medical imaging volumes are taken anisotropically with a high intra-slice resolution and a low inter-slice resolution. This phenomenon leads to problems such as rough or even broken tissue boundaries in 3D reconstructed models, which will undoubtedly a ect the accuracy of lesion analysis result. As such, an accurate and reliable method to upsample the low inter-slice resolution, which we refer to as the medical image slice interpolation techniques, is much needed in research. In addition to generating accurate 3D reconstructions, medical slice interpolation can also be widely used in medical image segmentation, multi-frame super-resolution (MFSR) reconstruction , and other elds. By adding new virtual slices between every two consecutive images, as shown in Fig. 1, the number and information of experimental data sets are increased, in order to boost MFSR and medical image segmentation accuracy. Especially, increasing the amount of training samples is indispensable for the popular research method such as neural network. Therefore, it is necessary to improve the inter-slice interpolation techniques to increase the axial spatial resolution of the data acquired using medical imaging modalities. Image interpolation technology has been wide-spread used in various area of image processing, especially in the eld of medical image processing. The methods for this task can be categorized into four groups. (1) Grayscale-based interpolation methods [12, 13] directly use the grayscale information of two consecutive images, to interpolate the inter-layer images through a set of basis functions. Nearest neighbor interpolation [2] , linear interpolation [1] and cubic B spline interpolation [11] are the common types of such interpolation methods. This method is widely used in image interpolation because of its computational simplicity and less computationally expensive. However, the interpolated images obtained by these methods are usually too smooth and contain the artifacts. (2) The shape-based interpolation methods [3, 7] generate contours of the image to be interpolated directly based on the contour shapes of two consecutive images. Compared with the grayscale-based interpolation methods, it can e ectively