Today, there are diseases of the human spine that require long treatment processes and are very difficult to detect. Detection of spinal disorders from computed tomography images is usually performed by our experienced physicians as a result of long efforts. The lack of personnel and the high number of cases in this regard cause the treatment process to be long. In this study, segmentation of the human spine with machine learning techniques is recommended to assist our physicians in disease detection. VerSe dataset was used for segmentation of IT backbones. VerSe data consists of 300 CT images. Wavelet transform and non-local averaging methods are used to improve the images in the VerSe dataset. The vertebrae of each case in this dataset are segmented with the UNet convolutional network adapted to this dataset. The images segmented with the UNet network were measured with the Dice Score and 99% accurate segmentation was achieved. Segmenting the vertebrae in VerSe computed tomography images with the UNet network with high success will ensure accurate detection of spinal diseases. As a result of image enhancement, restoration and segmentation, it is predicted that the treatment processes will be shortened, with higher quality images being read more accurately and faster. This study will enable the detection of spine specific disorders to be classified with a high success rate in the later stages.
(This study was produced from the doctoral thesis of the 1st ranked author).
Anahtar Kelimeler: image processing, neural networks, spine segmentation
|
