Abstract
In order to achieve an accurate attenuation correction in brain PET images acquired by hybrid PET/MR scanners, it is mandatory to delineate cortical bone and cavities in the MR images. Automated segmentation of the anatomical Ultra short echo time (UTE) MR images into different regions allows to assign them to the corresponding attenuation coefficients. The UTE sequence yields two components obtained by echo times TE=0.07 ms and TE=2.46 ms.
UTE images were first normalized by means of a scatterplot-based normalization technique, in which the scatterplot of a given scan is fitted into that of reference's. Second, a correction mask was generated to reduce the problem of the head edges resulting in the first component. Third, the fully automatic virtual extraction was realized by developing two methods: the two-class Support Vector Machine (C SVM) -based method and the single-class Support Vector Machine (S SVM)-based method using different kernels. Four datasets were evaluated with the corresponding registered CT scans and with an expert manual segmentation of the cavities.
The C SVM-based segmentation of the skull using the RBF kernel reached a Dice coefficient (D) value of 0.83±0.042 (mean ± SD). The S SVM-based segmentation of cavities using the RBF kernel attained a D value of 0.73±0.02. Based on the present results, the following conclusions can be drawn: First with our methods, the fully automatic segmentation of cortical bone and cavities reaches good results. Second, intensity normalization enables the development of the S SVMbased method for segmentation of cortical bone and cavities.
Keywords: Skull segmentation, UTE, Support vector machine, Attenuation correction, PET/MR, MRI.
Call for Papers in Thematic Issues
Intelligent Information Retrieval for Multispectral and Hyperspectral Imaging
Multispectral and hyperspectral imaging is an interdisciplinary research area, widely focuses on spectral, spatial, and temporal data. It offers a plethora of opportunities to efficiently analyze the vast area of the earth's surface. However, there are numerous factors such as dimensionality and size of the hyperspectral data, lesser training samples, ...read more
Machine and Deep Learning in Medical Image Diagnosis
Medical Image Processing regards a set of methodologies that have been developed over recent years with the purpose of improving medical image quality, improving medical data visualization, understanding, and assisting medical diagnosis, and so on. The application of machine and deep learning methods in sensing and imaging can potentially have ...read more
Quantum Machine Learning for Medical Data and Imaging
Quantum computing has promised a significant speedup in certain computationally intensive tasks that are intractable on classical computers. Researchers in quantum machine learning, such as those working in computer vision, image processing, biomedical analysis, and related topics, may play an important role in comprehending and working on complicated medical data, ...read more
4