Knowledge-based localization of hippocampus in human brain MRI

Abstract

We present a novel and efficient method for localization of human brain structures such as hippocampus. Landmark localization is important for segmentation and registration. This method follows a statistical roadmap, consisting of anatomical landmarks, to reach the desired structures. Using a set of desired and undesired landmarks, identified on a training set, we estimate Gaussian models and determine optimal search areas for desired landmarks. The statistical models form a set of rules to evaluate the extracted landmarks during the search procedure. When applied on 900 MR images of 10 epileptic patients, this method demonstrated an overall success rate of 83%.

Fig. 1

Hippocampus on T1-weighted images of brain MRI: a) coronal view with left hippocampus delineated where medial limits show no distinct boundaries, b) sagittal view with hippocampus delineated where anterior limits show no distinct boundaries, c) 3D model of hippocampus overlaid on coronal view.

Fig. 2

Generating binary images using FCM. a) Original image, b-d) binary images representing background and CSF, GM, WM, respectively, generated based on seven clusters produced by FCM, shown in (e)-(k). Note that segmented tissues are shown black in (b)-(k).

Fig. 3

An instance of marking points of interest: starting points of search: the most lateral points of lateral ventricles, desired points: superior landmarks of hippocampus, undesired points: medial landmarks of insular cortex.

Fig. 4

Marked landmarks and their estimated distributions for an example of six landmarks. a) Marked points with iso-countors drawn where only 5% of volume under estimated distribution function left outside. b) 3D plot of sum of estimated Gaussian distributions where undesired models are considered to be negative.

Fig. 5

a) Illustration of probability density functions (pdf) for 6 landmarks. pdf’s of desired landmarks are shown positive and those of undesired landmarks are shown negative. Black lines show the search angles. b) Zoomed-in view of the pdf’s intersecting with a search boundary.

Fig. 6

a) Schematic showing a case where distributions of landmarks on each individual slice/location are completely separate while their projections on 2D space overlap. b) An actual case illustration in which 3D iso-surface of desired and undesired landmarks are separate while their 2D projections overlap.

Fig. 7

a) Search method mimicking emission of radii of sight, b) fixed angle step-size and possibility of missing pixels while traversing search area. Squares present some of missing pixels, c) determining angle step-size adaptively leaves no pixel untouched/unexamined.

Fig. 8

Structuring elements of finding connected components (FCC) algorithm modified to grow directionally.

Fig. 9

Milestones of roadmap to localize left hippocampus: 1) starting point of roadmap, 2) most lateral landmark of left lateral ventricle, 3) a point on superior limits of left hippocampus, and 4) most inferior-medial point of left insular cortex. Solid arrows show the sequence in which milestones are visited.

Fig. 10

a) Absolute coordinates of lateral and medial points of lateral ventricles (pentagrams and triangles, respectively), superior, lateral, and inferior points of hippocampus (diamonds, stars, and pentagrams, respectively), and medial inferior points of insular cortex (triangles). All points are given in a coordinate system built on starting point of roadmap (define in Section 2.1) as origin. The iso-contours are drawn at 95% confidence level. (b)-(e) Deviations in y-coordinate for each pair of a landmark in left and right hemisphere vs. deviation of average of corresponding x-coordinates from brain midline, b) lateral, c) inferior landmarks of hippocampus, d) medial inferior points of insular cortex, e) lateral landmarks of lateral ventricles.

Fig. 11

Distribution of medial inferior landmarks of insular cortex shown as pentagram in coordinates system built on superior landmarks of hippocampus as origin.

Fig. 12

a)-d) Sagittal and coronal pairs of T1-weighted MRI’s of 2 subjects with initial contours overlaid. e)-h) Sagittal and coronal pairs of T1-weighted MRI’s with the final contours overlaid.

Fig. 12

a)-d) Sagittal and coronal pairs of T1-weighted MRI’s of 2 subjects with initial contours overlaid. e)-h) Sagittal and coronal pairs of T1-weighted MRI’s with the final contours overlaid.

Fig. 13

Sagittal cross sections of the hippocampus with initial polygons produced using 2 different sensitivity levels for initialization: a) 78% b) 41%. The final segmentation results (c) and (d) produced by deformable model using initial models shown in (a) and (b), respectively. Note robustness of proposed method.

Fig. 14

a), b) Coronal and sagittal views of a base MR volumetric data with the hippocampus boundaries overlaid. c), d) Hippocampus boundaries transferred from the base to a new MRI dataset using mutual information registration. e) Coronal view of segmented model produced by deformable model using initial model shown in (c) and (d): note an absolute failure. f) a superior point of the hippocampus transferred model is used to search for the lateral ventricles based on statistical models estimated in Section 2.1. g) Landmarks of interest found by initialization method. h) Result of deformable model applied on landmarks shown in (g).

Fig. 15

a) BK&CSF binary image on which lateral and superior limits of head are determined to find starting point of search. b) Manual landmark identification for lateral ventricles when looking from starting point (point-1); lateral ventricles (point-2) as desired landmark, top corners of third ventricle (point-3), Sylvian fissure (point-4), and lateral points of callosal sulcus (point-5) and eingulate sulcus (point-6) as undesired landmarks. c) Distribution of desired and undesired landmark points with iso-contours drawn at 95% confidence level.

Fig. 16

a) Manual landmark identification for desired and undesired structures when looking for superior points of hippocampus from lateral points of lateral ventricles (point-1). Medial limits of insular cortex (point-2) and lateral inferior limits of hypothalamus (point-3) are undesired landmarks, and superior point of hippocampus (point-4) is desired landmarks. b) Manually marked desired and undesired landmarks with iso-contours at 95% confidence level. Search area for left hemisphere is depicted. Note that viewpoint is moved to an optimal position (marked by an arrow on (b)) to balance sensitivity and specificity.

Fig. 17

a) Manual landmark identification for desired and undesired structures when looking from superior points of hippocampus (point-4). Medial point of insular cortex (point-2) is desired point. Lateral limits of hypothalamus (point-1) and medial limits of SMGTI (point-3) are undesired landmarks. b) Search performed on a GM binary image to find a black point with at least six neighboring black points in its 8-nearest neighborhood. c) Region connected to point found in (b) is grown downward by applying FCC.

Fig. 18

a) Manual landmark identification for desired and undesired structures when looking from medial inferior point of insular cortex (point-1) for lateral points of hippocampus (point-2) as desired landmarks. Medial limits of SMGTI (point-4), and superior limits of PFGI (point-3) are undesired landmarks. b) Manual landmark identification when looking for inferior point of hippocampus (point-1) as desired landmark from superior point of this structure (point-2). Lateral limits of PFGI (point-3), and lateral boundaries of peduncular (point-4) are marked as undesired landmarks. c) Distribution of desired and undesired landmarks with iso-contour drawn at 95% confidence level for landmarks marked on (b). Arrow and letter “P” mark optimal viewpoint.