Qingmao Hu

On minimum variance thresholding

Variance-based thresholding methods could be biased from the threshold found by expert and the underlying mechanism responsible for this bias is explored in this paper. An analysis on the minimum class variance thresholding (MCVT) and the Otsu method, which minimizes the within-class variance, is carried out. It turns out that the bias for the Otsu method is due to differences in class variances or class probabilities and the resulting threshold is biased towards the component with larger class variance or larger class probability. The MCVT method is found to be similar to the minimum error thresholding.

A rapid algorithm for robust and automatic extraction of the midsagittal plane of the human cerebrum from neuroimages based on local symmetry and outlier removal

A rapid algorithm for robust, accurate, and automatic extraction of the midsagittal plane (MSP) of the human cerebrum from normal and pathological neuroimages is proposed. The MSP is defined as a plane formed from the interhemispheric fissure line segments having the dominant orientation. The algorithm extracts the MSP in four steps: (1) determine suitable axial slices for processing, (2) localize the fissure line segments on them, (3) select inliers from the extracted fissure line segments through histogram-based outlier removal, and (4) calculate the equation of the MSP from the selected inliers. The fissure line segments are localized by minimizing the local symmetry index characterizing anatomical properties of images in the vicinity of the interhemispheric fissure. A two-stage angular and distance outlier removal is introduced to handle abnormalities. The algorithm has been validated quantitatively with 125 structural MRI and CT cases from 10 centers on three continents by studying its accuracy; tolerance to rotation, noise, asymmetry, and bias field; sensitivity to parameters; and performance. A statistical relationship between algorithm accuracy and the datas adherence to planarity is also determined. The algorithm extracts the MSP below 6 s on Pentium 4 (2.4 GHz) with the average angular and distance errors of (0.40 degrees; 0.63 mm) for normal and (0.59 degrees; 0.73 mm) for pathological cases. The robustness to noise, asymmetry, rotation, and bias field is achieved by extracting the MSP based on the dominant orientation and local symmetry index. A low computational cost results from applying simple operations capturing intrinsic anatomic features, constraining the searching space to the local vicinity of the interhemispheric fissure, and formulating a noniterative algorithm with a coarse and fine fixed-step searching. In comparison to the existing methods, our algorithm is much faster, performs accurately and robustly for a wide range of diversified data, and is fully automatic and thoroughly validated, which make it suitable for clinical applications.

Thresholding based on variance and intensity contrast

A new thresholding criterion is formulated for segmenting small objects by exploring the knowledge about intensity contrast. It is the weighted sum of within-class variance and intensity contrast between the object and background. Theoretical bounds of the weight are given for the uniformly distributed background and object, followed by the procedure to estimate the weight from prior knowledge. Tests against two real and two synthetic images show that small objects can be extracted successfully irrespective of the complexity of background and difference in class sizes.

Supervised range-constrained thresholding

A novel thresholding approach to confine the intensity frequency range of the object based on supervision is introduced. It consists of three steps. First, the region of interest (ROI) is determined in the image. Then, from the histogram of the ROI, the frequency range in which the proportion of the background to the ROI varies is estimated through supervision. Finally, the threshold is determined by minimizing the classification error within the constrained variable background range. The performance of the approach has been validated against 54 brain MR images, including images with severe intensity inhomogeneity and/or noise, CT chest images, and the Cameraman image. Compared with nonsupervised thresholding methods, the proposed approach is substantially more robust and more reliable. It is also computationally efficient and can be applied to a wide class of computer vision problems, such as character recognition, fingerprint identification, and segmentation of a wide variety of medical images.

Fast connected-component labelling in three-dimensional binary images based on iterative recursion

We propose two new methods to label connected components based on iterative recursion: one directly labels an original binary image while the other labels the boundary voxels followed by one-pass labelling of non-boundary object voxels. The novelty of the proposed methods is a fast labelling of large datasets without stack overflow and a flexible trade-off between speed and memory. For each iterative recursion: (1) the original volume is scanned in the raster order and an initially unlabelled object voxel v is selected, (2) a sub-volume with a user-defined size is formed around the selected voxel v, (3) within this sub-volume all object voxels 26-connected to v are labelled using iterations; and (4) subsequent iterative recursions are initiated from those border object voxels of the sub-volume that are 26-connected to v. Our experiments show that the time-memory trade-off is that the decrease in the execution time by one-third requires the increase in memory size by 3 orders. This trade-off is controlled by the user by changing the size of the sub-volume. Experiments on large three-dimensional brain phantom datasets (362x432x362 voxels of 56 MB (megabytes)) show that the proposed methods are three times faster on the average (with the maximum speedup of 10) than the existing iterative methods based on label equivalences with less than 1 MB memory consumption. Moreover, our algorithms are applicable to any dimensional data and are less dependant on the geometric complexity of connected components.

Fast Talairach Transformation for Magnetic Resonance Neuroimages

Abstract: We introduce and validate the Fast Talairach Transformation (FTT). FTT is a rapid version of the Talairach transformation (TT) with the modified Talairach landmarks. Landmark identification is fully automatic and done in 3 steps: calculation of midsagittal plane, computing of anterior commissure (AC) and posterior commissure (PC) landmarks, and calculation of cortical landmarks. To perform these steps, we use fast and anatomy-based algorithms employing simple operations. FTT was validated for 215 diversified T1-weighted and spoiled gradient recalled (SPGR) MRI data sets. It calculates the landmarks and warps the Talairach-Tournoux atlas fully automatically in about 5 sec on a standard computer. The average distance errors in landmark localization are (in mm): 1.16 (AC), 1.49 (PC), 0.08 (left), 0.13 (right), 0.48 (anterior), 0.16 (posterior), 0.35 (superior), and 0.52 (inferior). Extensions to FTT by introducing additional landmarks and applying nonlinear warping against the ventricular system are addressed. Application of FTT to other brain atlases of anatomy, function, tracts, cerebrovasculature, and blood supply territories is discussed. FTT may be useful in a clinical setting and research environment: (1) when the TT is used traditionally, (2) when a global brain structure positioning with quick searching and labeling is required, (3) in urgent cases for quick image interpretation (eg, acute stroke), (4) when the difference between nonlinear and piecewise linear warping is negligible, (5) when automatic processing of a large number of cases is required, (6) as an initial atlas-scan alignment before performing nonlinear warping, and (7) as an initial atlas-guided segmentation of brain structures before further local processing.

A knowledge-driven algorithm for a rapid and automatic extraction of the human cerebral ventricular system from MR neuroimages.

A knowledge-driven algorithm for a rapid, robust, accurate, and automatic extraction of the human cerebral ventricular system from MR neuroimages is proposed. Its novelty is in combination of neuroanatomy, radiological properties, and variability of the ventricular system with image processing techniques. The ventricular system is divided into six 3D regions: bodies and inferior horns of the lateral ventricles, third ventricle, and fourth ventricle. Within each ventricular region, a 2D region of interest (ROI) is defined based on anatomy and variability. Each ventricular region is further subdivided into subregions, and conditions detecting and preventing leakage into the extra-ventricular space are specified for each subregion. The algorithm extracts the ventricular system by (1) processing each ROI (to calculate its local statistics, determine local intensity ranges of cerebrospinal fluid and gray and white matters, set a seed point within the ROI, grow region directionally in 3D, check anti-leakage conditions, and correct growing if leakage occurred) and (2) connecting all unconnected regions grown by relaxing growing conditions. The algorithm was validated qualitatively on 68 and quantitatively on 38 MRI normal and pathological cases (30 clinical, 20 MGH Brain Repository, and 18 MNI BrainWeb data sets). It runs successfully for normal and pathological cases provided that the slice thickness is less than 3.0 mm in axial and less than 2.0 mm in coronal directions, and the data do not have a high inter-slice intensity variability. The algorithm also works satisfactorily in the presence of up to 9% noise and up to 40% RF inhomogeneity for the BrainWeb data. The running time is less than 5 s on a Pentium 4, 2.0 GHz PC. The best overlap metric between the results of a radiology expert and the algorithm is 0.9879 and the worst 0.9527; the mean and standard deviation of the overlap metric are 0.9723 and 0.01087, respectively.

Segmentation of brain from computed tomography head images

An algorithm to determine the human brain (gray matter (GM) and white matter (WM)) from computed tomography (CT) head volumes with large slice thickness is proposed based on thresholding and brain mask propagation. Firstly, a 2D reference image is chosen to represent the intensity characteristics of the original 3D data set. Secondly, the region of interest of the reference image is determined as the space enclosed by the skull. Fuzzy C-means clustering is employed to determine the threshold for head mask and the low threshold for brain segmentation. The high threshold is calculated as the weighted intensity average of the boundary pixels between bones and GM/WM. Based on the low and high thresholds, the CT volume is binarized, followed by finding the brain candidates through distance criterion. Finally the brain is identified through brain mask propagation using the spatial relationship of neighboring axial slices. The algorithm has been validated against one non-enhanced CT and one enhanced CT volume with pathology

Fast, accurate, and automatic extraction of the modified Talairach cortical landmarks from magnetic resonance images.

The Talairach transformation is the most prevalent way to normalize brains and is hindered by, among others things, a lack of automatic determination of cortical landmarks. An algorithm to locate the modified Talairach cortical landmarks in three steps is proposed: determination of the three planes containing the landmarks; segmentation of the planes based on range‐constrained thresholding and morphologic operations; and local refinement of the segmentation to locate the landmarks. The algorithm has been validated against 62 T1‐weighted and SPGR MR diversified data sets. For each data set, it takes less than 2 s on a Pentium 4 to extract all six landmarks. The average landmark location errors are below 0.9 mm. The algorithm is robust due to incorporation of anatomic knowledge. A low computational cost results from processing of three 2D images and employing only simple operations like thresholding, basic morphologic operations, and distance transform. Magn Reson Med 53:970–976, 2005.

Regularized fuzzy c-means method for brain tissue clustering

This paper presents a regularized fuzzy c-means clustering method for brain tissue segmentation from magnetic resonance images. A regularizer of the total variation type is explored and a method to estimate the regularization parameter is proposed.