Image Data

Rat MRI scans

Goals

Water and fat decomposition and analysis in rat body

Methods I, IP

Graph-cut, Optimal surface detection

Methods II, AI

Use three-point Dixon decomposition model and harmonic retrieval framework to estimate fieldmap; Build up graph model with new cost function and smooth constraint to recover fat and water signal respectively

Validation

Method will be tested on 2~3 other MRI datasets. Comparisons will be made with IDEAL and other graph based decomposition methods. Compare in a meaningful voxel-based and quantitative way with existing competing methods. Also, utilize the phantom images to assess the correctness of your decomposition.

Image Data

Surface extraction of human trabecular bone from 3D CT data of trabecular bone

Goals

Develop an improved sub-voxel meshing approach for trabecular bone structure representation.

Methods I, IP

Extract bone surfaces from original CT volume data in sub-voxel resolution, and represent as a mesh, based on bone density information.

Methods II, AI

Use knowledge of feasible 3D geometry of the trabecular bone to influence the meshing process so that the resulting meshes are legal, and anatomically feasible.

Validation

In 15 CT datasets, compare with marching cubes and dual contour methods - determine a set of quantitative indices that are relevant for determining the quality of the resulting mesh.

Image Data

Establish the tree structure of the arteries and veins and segment arteries and veins in fundus images

Goals

2D fundus color images, with ground truth of manually segmented arteries and veins

Methods I, IP

The method is to combine the tracking-based method and the supervised method. The tracking-based method is to obtain the connectivity between neighboring vessel segments in terms of possibility.

The supervised method is to classify the possibility of vessel segments to be arteries or veins. In the supervised method I will use a new feature descriptor and compare different classifiers (SVM, NN, kNN, and maybe random forest) to obtain the best result of probability map.

The separation of arteries and veins is transferred to an optimization problem in which the total possibility (The possibility of vessel segments and the possibility of vessel connectivity) is maximized. The optimization problem can be solved using some stochastic search algorithms like TABU search, or genetic algorithm to approach a near-optimal solution.

Use tracking-based method to find the centerlines of vessel segments. And use the morphological information of vessel segments to obtain the probability of connectivity of neighboring vessel segments.

Methods II, AI

Use supervised method to segment vessel segments. Then use classification method to obtain the probability of every vessel being artery or vein.

Validation

Compared the results in at least 30 2-D color fundus datasets with the ground truth of manually separated arteries and veins. Pixel-wise accuracy and vessel-wise (segment-focused assessment of performance) accuracy are calculated.

Image Data

3D CT bone images - trabecular bone

Goals

Improve the binary skeleton pruning algorithm so that pruning primarily removes unwanted skeleton segments.

Methods I, IP

First, use the existing 3D thinning approach to determine the skeleton of binary object.
Second, develop improvements to the existing digital topology analysis (DTA) method to classify each voxel's topological type. Next, use the DTA results to develop a method for segmenting different branches of the skeleton image.

Methods II, AI

Distinguish true branches and noisy/spurious branches in the skeleton image, possibly using statistical analysis methods to explore the geometric and topological features/properties. Utilize both global and local information. Develop a decision making algorithm to identify and remove noisy/spurious branches according to pre-specified requrements for relevant branch characteristics.

Validation

15 CT images of trabecular bone and sinusoidal cross synthetic phantoms to test performance. Compare the results with main existing popular skeleton/skeleton pruning algorithm. Perform quantitative validation using a properly defined independednt standard.

Image Data

3D micro CT image of newborn pig's excised lung specimen, 5 different inflation pressures

Goals

Newborn pig's lung segmentation and registration to measure how the image has changed when it is inflated

Methods I, IP

Intensity based segmentation of parenchymal surface, B-Spline registration across inflation phases, region growing to yield air/tissue segmentation

Methods II, AI

Use fuzzy connectivity method to segment lung region. Use minimization method to minimize registration error - sort of a 3-step or multi-step register/segment/re-register. The knowledge based part wil be to learn/determine what should be used as objective function for the minimization.

Validation

To validate segmentation method, compare with manually segmented lung and calculate ROI volume. Using Jacobian matrix to validate registration result. Use at least 3 sets of 5-inflation pressure image datasets.

Image Data

3D SD-OCT scans (Zeiss Cirrus)

Goals

Segment Choroid Vessels in 3D, measure Choroid layer thickness and track the tree-structure of  Choroid Vessels

Methods I, IP

Vessel Detection – multi-scale Hessian Analysis; Vessel Segmentation – Region Growing (or Graph Cut); Vessel Central Axis Detection – 3D Skeletonization

Methods II, AI

Use Thin Plate Spline to fit the top/bottom envelope surfaces of the Choroid layer and start with the end-points of the vessel central axis to back-track the separate sub-trees from the segmentation results - smartness/a priori knowledge will be needed to separate the "graph" to "trees".

(If this task of subtree separation would not be feasible, explore the ability of this approach to identify choroidal SEAD footprint locations, or possibly choroidal SEAD locations in dry AMD.)

Validation

Manual tracings or - if possible, compare with ICG angiographic projection images, find out whether it is possible to do ICG for arterial or venous filling phase only. Develop a meaningful set of quantitativbe indices demonstrating succes of your method.

Image Data

fMRI scans (BOLD or msMRI = magnetic source MRI), 3D human brain

Goals

Find an alternate method to identify activations

Methods I, IP

Use alternate segmentation methods, as current methods are essentially binary/threshold methods

Methods II, AI

Replace Cross Correlation analysis with a knowledge-based method, perhaps ICA

Validation

In 10 BOLD-msMRI pairs, compare results with existing Cross Correlation analysis methods currently in use on BOLD vs. msMRI data. use DICE or similar considering locations and activation region sizes.

Image Data

3D CT bone image - human tibia

Goals

The purpose is to segment cortical bone out from bone marrow in 3D CT bone image

Methods I, IP

graph-based method - simultaneous dual surface

Methods II, AI

automated design of cost functin from examples of marrow and bone, consider using regional AND surface based cost function

Validation

25 CT images, independent standard needs to be defined, use leave X out approach to train the cost function and determine surface positioning and/or DICE performance.

Image Data

3D Sagittal Dual Echo Steady State (DESS) Knee MRI from Osteoarthritis Initiative (OAI).

Goals

Develop an automated technique to segment and quantitatively analyze the sub regions of the cartilage surface to study loss due to onset of Osteoarthritis.

Methods I, IP

LOGISMOS segmentation of cartilage/bone in the knee joint, identification of the proper landmarks, fitting a sphere, definining the proper regions on the femur and tibia. Consider incorporating 3D interactive edit/resegment approach to obtain the best segmentation possible - this part is part of you "other" research.

Methods II, AI

Knowledge of the femoral anatomy will be used for identification of the proper landmarks as described above.

Validation

The quantitative analysis of the sub regions of the manually segmented knees will be performed in 40 knees, assessment based on comparison with Eckstein-provided cartilage measurements.

Independently, the accuracy of landmar detection, cutting-plane definition, and fitted sphere will be performed - comparison with expert detection of these landmarks and parameters.

Image Data

optic-nerve-head centered SD-OCT volumes, 3D, monkey eyes

Goals

segment the anterior surface of the lamina cribrosa (where visible); also use GPU as part of approach, if possible, but not as requirement

Methods II, IP

graph-based approach for finding final surface

Methods I, AI

use a priori anatomical knowledge to help determine approximate location of final surface (e.g., use of vessel positions, use of neural canal opening location); use classification approach for learning appropriate cost functions

Validation

use SD-OCT volumes from 10 monkey eyes; use leave-x% out to train/test (with respect to cost-function design); compare with manual tracings; if using GPU, also compare results and running time with/without use of GPU

Image Data

HD 5-line OCT raster scans and SD-OCT volumes from subjects with swollen optic disc (or called papilledema); at least 15 pairs with various degrees of optic disc swelling

Goals

To use HD 5-line raster scan to aid the layer segmentations in SD-OCT scans with papilledema. By using the information from these HD scans, the segmentation of retinal nerve fiber layer (RNFL) would be improved in the volumetric scans

Methods I, IP

First, to register these HD 5-line raster scans within volumetric scans, and then to use graph-based search to perform the layer segmentations.

Methods II, AI

Using a priori knowledge of the pattern of HD 5-line raster scans to help initial placement within SD-OCT volumetric scans, and using a priori knowledge of swollen optic-nerve-head (ONH) shapes to help specify cost functions as well as constraints in the graph-based search algorithm.

Validation

First, comparing automated registration with that from manual placement, and then comparing the segmentation results with manual tracing (on subset of slices)

Image Data

IVUS datasets , baseline and follow up, 25 pairs of 3D datasets

Goals

identify frame-to-frame and wedge-to-wedge differences based morphologic and plaque characterization and TCFA indices between the correspondng locations in baseline anf follow up data, identify locations of expected change in FU data from BL data.

Methods I, IP

Improve existing registration of BL-FU IVUS pullbacks so that there is frame-to-frame match and wedge-to-wedge match for all frames in the BL/FU datasets, interpolation may be needed. Recalculate all indices accordingly. Develop several promising not-so-far thought-of indices of morphology/vulnerability/thin plaque cap/etc.

Methods II, AI

Using R and single/multivariate regression analysis, identifylocations of sufficient change between BL and FU for frames/wedges, maybe regions.Attempt to identify locations that will change from BL data only

Validation

Percent success rate of identification of locations that will change.

Image Data

34 SD-OCT volumes + stereo fundus image pairs of the ONH (from subjects with glaucoma or suspicion of glaucoma)

Goals

Segment the optic disc/cup simultaneously from both modalities

Methods I, IP

Graph-based approach from finding disc/cup boundaries simultaneously (across both modalities)

Methods II, AI

Use classification approach (such as kNN) for learning appropriate cost functions

Validation

Use leave x% out approach to train/test (using manual tracings from experts); also compare to existing single modality approaches - validation based on both the segmentation errors and on comparison of cup/disc areas and cup/disc ratios.

Image Data

3D CT data of fractured and intact ankle bones.

Goals

Segment cortical and cancellous bone fragments from highly comminuted tibial plafond fractures. Following segmentation classify fragment surfaces as periosteal, articular, or cancellous.

Methods I, IP

Multi-object fuzzy connectivity  segmentation

Methods II, AI

A priori knowledge about  location of tibia, fibula and talus in CT images. Knowledge about curvature and relative intensity values of periosteal surface vs cancellous and articular

Validation

Compare against previously hand segmented CT scans to determine segmentation accuracy.  Compare against previously hand classified segmentations to determine classification accuracy.

Image Data

Liver CT scans, human 3D CT

Goals

Segment liver tumors in 3D

Methods I, IP

Detect the tumor location = detect the star centers automatically.

Methods II, AI

Maximum weight Digital Regions Decomposable into Digital Star-Shaped Regions

Validation

8 (better 15 if possible, talk to Prof. Beichel) liver CT scans with liver tumor will be segmented using the proposed approach. Segmentation results will be compared with independent standard to get quantitative analysis results including DICE coefficients, etc.

Image Data

Multi-spectral images from IR and low-light visible sensors - in movie sequences

Goals

Multi-spectral fusion - Come up with an intelligent way of determining the optimal/good weighting scheme for combining the 2 sources

Methods I, IP

Develop a weighting scheme so that - if the weights are known - two images can be fused, e.g., after having been represented using a wavelet based method, several approaches will be implemented and tested

Methods II, AI

Develop a knowledge based approach that allows to combine the two wavelet based image source representations (some sort of fuzzy logic or ANN) so that the outcome is optimal wrt some meaningful objective function. A self-learning approach would be cool.

Validation

In 10 sequences with at least 20 frames each - determine the quality of the fusion in comparison ot several (5) suboptimal fusions, demonstrate tha the optimal fusion is perceived better than the suboptimal fusions wrt to some usefulness criterion.