Image Data

OCT images of retina 2D+3D, RTA images in 3D, (possibly 2D+T carotid ultrasound images)

Goals

Develop methods for retinal analysis

Methods I, IP

3D graph search, use of border + region information in cost functions

Methods II, AI

Incorporate a priori knowledge about pathologic and normal patterns in the image analysis algorithms

Validation

30 OCT and/or RTA images, compare to expert-defined independent standard

Image Data

Multimodality data sets including MRI, pet/CT of human brain/mouse images from Biomedical Imaging Research Lab at USC.

Goals

Develop a non-rigid mutual information registration algorithm parameterized on fourier basis. Use a priori image-based information for design of the mutual registration objective function.

Methods I, IP

Use existing methods to write registration cost function based on mutual information theory. Parameterize the cost function based on Fourier basis. Hierarchical approach will be used.

Methods II, AI

To increase the statistical power of the undergoing joint probability, the key factor for accuracy and speed, combine a prior joint probability which is computed from a pre-registered training set.

Validation

Simple phantom data sets with known transformations will be used
to verify the algorithm. Also compute relative overlap measurements
after applying this non-rigid MI registration algorithm, and
after rigid MI registration, compare the results. Validation will be performed in at least 15 3D data sets.

Image Data

Mouse micro CT data sets in 3D

Goals

Segment lungs, airways, vascular trees, concentrate on performance in lung segmentation, the airways and vessels are add-ons

Methods I, IP

Model-based lung segmentation, using the knowledge derived below

Methods II, AI

Derive knowledge about lung shapes in 3D from the independent standard

Validation

Validate in at least 15 micro CT data sets in comparison with expert-defined independent standard, using leave one out approach

Image Data

MR aortic images in 3D - from connective tissue disorder patients and from normals

Goals

Detect the differences of aorta 3D shape and 4D function between normal volunteers and tissue disorder subjects

Methods I, IP

Level set, cylindrical surface graph search method

Methods II, AI

Use Neural Networks or Support vector machines to classify the samples (normal samples, tissue disorder samples)

Validation

30 aortic images - validate against normal/patient status, and using disease severity rating as assessed by expert physician diagnosis

Image Data

Limited FOV human CT lung images - 15 time points over several breathing cycles

Goals

Track corresponding lung locations throughout the breathing cycle using the limited FOV data

Methods I, IP

Image registration and automated segmentation to track moving object.

Methods II, AI

Model of a breathing lung exists, and will be incorporated in the 2nd stage registration process

Validation

Validation in at least 10 subjects - validation will be performed in comparison with the independent standard.

Image Data

Multimodality data sets including MRI, PET, CT of human and/or sheep lung images

Goals

Define lobar and airway segmentation on a given lung image volume using inter-modality image registration from a segmented reference image (atlas).

Methods I, IP

inter-modality image registration with B-spline basis to find a mapping between the reference image volume and the given dataset.

Methods II, AI

Use the transformation function obtained above to deform the segmentations from the reference image to the candidate dataset, thereby defining segmentations on the candidate dataset. Incorporate knowledge about the potential different airway topologies to automatically identify the proper atlas mutation.

Validation

Quantitatively compare the results using independently segmented images as the reference volumes. Assess correctness wrt independent standard in at least 15 multi-modality data sets.

Image Data

3D CT images of sheep lungs imaged at 5 lung volumes

Goals

Identify point-to-point correspondences in expanding lungs using issue-property based model of lung expansion as a function of pressure

Methods I, IP

Segment lungs, lobes, airways, vessels using existing techniques. Using the model (below), achieve point-to-point registration of the lung parenchyma, airways, vessels, etc.

Methods II, AI

Partition lung in subvolumes of homogeneous tissue properties with estimated (literature based) mechanical properties. Use it to build a model of pressure-expansion relationships

Validation

Using implanted beads and other identifiable landmarks in 8 sheep data sets with 65 lung volumes.

Image Data

Stereo images of retinal vessels

Goals

Identify correspondence between stereo views, branch points, segments, etc. Resolve branching/overlapping structures in the stereo views. Determine vessel diameter for a specific segment

Methods I, IP

Region & edge based identification of vessel trees, epipolar geometry to determine stereo correspondence, generate a formal tree/graph structure, identify "conflicts" = closed loops in the graphs.

Methods II, AI

Bifurcation model based + stereo based resolution of conflicts in the graph = resolve which is bifurcation and which is overlap.

Validation

In at least 20 stereo pairs, against expert analysis. Validate both diameters and bifurcations/overlaps.

Image Data

3D pulmonary CT data from mild and severe emphysema subjects

Goals

Lung lobe segmentation (all 3 fissures)

Methods I, IP

3D surface graph search segmentation guided by zone of influence from vessel + airway segmentation

Methods II, AI

A priori knowledge about fissure shapes and locations shall be incorporated to overcome lack of image information due to emphysema tissue disintegration

Validation

15 emphysema data sets, comparison against expert tracing in selected slices

Image Data

Retinal optic disc images

Goals

Automate optic disc grading. (ophthalmology image analysis)

Methods I, IP

Optic nerve segmentation, detection of cup/disc ratio - utilize stereo epipolar geometry if possible. Consider using 3D graph search for simultaneous detection of 4 borders from a single stereo pair.

Methods II, AI

Classification based cup/disc segmentation. Consider also combining the border based and classification segmentation in a single process.

Validation

Validate against independent standard, as well as compare the segmentation based, the classification based and combined approaches with each other. Validate in the available 15 stereo pairs.

Image Data

Optic disc stereo pairs

Goals

Automate optic disc grading

Methods I, IP

Apply edge-based together with region based and shape based segmentation to find the borders of cup, and disc. Utilize the segmented vessel info to improve the outcome.

Methods II, AI

Determine the border between of cup, disc, and background that are occluded by blood vessels. Utilize stereo pair information to increase robustness of the analysis.

Validation

Compare the ratio of cup to disc by automate optic grading with that given by ophthalmologists. Compare results obtained without and with utilization of the stereo information. Validate in the available 15 stereo pairs.

Image Data

Human CT images of normals and emphysema patients

Goals

Develop a method for identification of abnormally shape rib cage

Methods I, IP

Segment ribs in 3D, label/name individual ribs, identify reproducible landmarks that can be used for building a pont distribution model.

Methods II, AI

Build a PDM of the rib cage population and use for classification of subjects in normal/abnormal categories.

Validation

Validate classification results against patient disease status information, use leave one out approach - validate in at least 10 normal and 10 emphysema patients.

Image Data

MRI diffusion tensor brain data

Goals

Develop new fiber tracking algorithm capable of crossing locally ambiguous data areas, develop a method for distinguishing between normal and abnormal anatomy

Methods I, IP

Apply 3D graph search method on fiber tracking, fiber clustering, spatial fiber co-registration and comparison.

Methods II, AI

Apply SVM for fiber classification. Study the white matter group difference between patient and normal controls.

Validation

16 schizophrenics and normals - determine classification correctness wrt patient diagnosis. Compare classification outcome when scan 1 or scan2 are used separately, and when both scans are used in a combined fashion.

Image Data

3D LIMA, 3D microCT, histology

Goals

Develop a method for identification of tissue in LIMA and/or microCT images based on matched histology – determine the following structures: airways, airway walls, vasculature, alveolar tissue (parenchyma), and determine regions of inflammation in any of these tissue classes.

Methods I, IP

Identification of vessels and airways in 3D sets of 2D LIMA and/or microCT images, formation of 3D vascular and airway objects, registration of LIMA/microCT and histology images

Methods II, AI

Develop a classification method for automated identification of inflammation regions on the LIMA/microCT images – train and test using the registered histology-based independent standard.

Validation

10 samples (3D cubes) of lung tissue, their LIMA and microCT images, plus associated histology, validation using a leave one out approach.

Image Data

3D MR images of ankle and knee.

Goals

Develop a method for robust detection of bones in 3D MR images.

Methods I, IP

Use level-set based or region-growing-based approach.

Methods II, AI

Using a standard 3D anatomy of the ankle/knee, develop automated initialization of the level set techniques so that the level set segmentation is automatically initialized. Incorporate a 3D graph representing the joint anatomy to represent the anatomical knowledge. The method must be robust with respect to normal anatomical variability, left/right extremities, coronal/sagittal scans, etc.

Validation

10-15 MR images of joints, in DICOM and Analyze formats, against manually-traced independent standard.