Advanced Techniques in Medical Imaging: A Comparative Study of Dimensionality Reduction Methods

Read full text Summarise with VIRAT
Authors: K. Vijay Anand, R. Srinivas Yadav, G. Rama Krishna
Cited by: 0

Summary

No summary yet. VIRAT can read this paper and write a plain-language summary of its aim, methods, key findings and limitations.

Summarise with VIRAT Free to try with your Science Catalogs account.

Abstract

This Paper represents a method for the automatic segmentation of the brain in magnetic resonance (MR) images of the human head. The method identifies brain areas of interest, including the gyri and other subcortical structures that were manually delineated in a set of labelled training images. Principal components analysis (PCA) is applied to the training ensemble in order to learn a PCA atlas subspace, which is a dimensionality-reduced linear subspace of labelled brain images. We employ this subspace to segment and label previously unseen subject images. This is accomplished by finding the PCA atlas closest to an input subject image through orthogonal projection of the latter into the subspace. The PCA atlas is then non rigidly registered to the subject image and the non rigid transformation is used to transfer the labels from the former to the latter, thereby segmenting the subject image. Our method is compared with alternative methods and the results are validated using overlap and distance metrics. KEYWORDS: MRI Images, Principal Component Analysis Atlas. The segmentation of brain images involves localizing and delineating the brain's different anatomical structures within three-dimensional (3D) tomographic images of the head. Brain image segmentation is important for longitudinal studies investigating how medications and other interventions affect the shapes and volumes of brain tissues, for surgical planning and for diagnosing disease. Traditionally, the various parts of the brain are segmented by trained experts who manually identify and delineate regions of interest. This process can be very Difficult and time consuming. It requires the design of complex protocols for identifying brain structures. The results from different manual segmentations using the same protocol may be inconsistent, which raises issues if a set of images in a study are delineated by different individuals. With the advent of non-invasive medical imaging devices that can easily acquire large quantities of tomographic data, manual segmentation becomes impractical for many studies. An alternative to manual segmentation is to develop computer algorithms that can automatically segment anatomical structures of interest in images of the brain. The automatic 3D brain image segmentation algorithm developed in this work involves computing a 3D atlas consisting of labelled voxels, from among a space of possible atlases learned from an ensemble of manually segmented training images, that is most suitable for segmenting an unlabeled subject image. This optimal atlas is then non rigidly registered to the subject image. Finally, the non rigid transformation is used to transfer the labels from the atlas to the subject image, thus segmenting the subject image. In greater detail, our procedure is as follows: Registration is the process of establishing correspondences between two images by finding a transformation of a source image into the coordinate system of a target image. First a set of training data comprising brain intensity images and their corresponding labels are mutually registered to minimize the variance across the data. This ensures that the variation in the data is attributed mostly to the differences in the shapes and textures of the anatomical structures of the brain and not to their locations, orientations, and sizes. Principal components analysis (PCA) is then performed on the intensity and label training data to learn a dimensionally-reduced linear subspace that is capable of generating labelled images that interpolate the training data. We refer to this subspace as a PCA atlas subspace. Then, an orthogonal projection is performed to find the point in the PCA atlas subspace that is closest Google Scholar : https://scholar.google.com/scholar?q=Advanced+Techniques+in+Medical+Imaging%3A+A+Comparative+Study+of+Dimensionality+Reduction+Methods DOI : https://doi.org/10.5281/zenodo.19223104 CITE PAPER Google Scholar

Fields of Study & Subjects

Medical Image Segmentation Techniques Brain Tumor Detection and Classification Advanced MRI Techniques and Applications

Publication details

Journal / source
Zenodo (CERN European Organization for Nuclear Research)
Publication date
30 January 2026
Language
EN