Image Data

Multi-slice CT of human lungs, volumetric, at least 30 volumes; get independent standard from Dr. McLennan et al. Contact Hidenori whether he has some data sets or nodule phantoms.

Goals

Detect lung nodules in 3D

Methods I, IP

In addition to raw image data, use image-based features as additional input to CNN.

Methods II, AI

Use Cellular Neural Nets - train/test on separate data, possibly use leave-one-subject-out

Validation

Leave-one-out, determine classification correctness in % - in comparison with human-defined independent standard. First validation on computer phantom.

Image Data

3-D digital Doppler ultrasound and phase velocity cine MRI images of flow through cardiac valves.

Goals

Automated segmentation of cardiac valves and measurement of flow integral during cardiac cycle.

Methods I, IP

Dynamic programming for segmentation of PVC MRI images - first individual frames in 2D - possibly a 3D as next step.

Methods II, AI

Active Appearance Model for segmentation of ultrasound images, again, maybe DP will be sufficient. Include a way how to identify the valve areas automatically (HT?).

Validation

Use expert traced PVC MRI and hand-corrected segmented PVC MRI images. Use expert traced ultrasound images as training set. Validate by comparing normalized flow rate measurements of ultrasound and PVC MRI. Start validation in computer phantoms.

Image Data

3D CT Lung Datasets, 5 with emphysema, 5 normal - more if available.

Goals

Develop a 3D texture classification program using a Support Vector Machine

Methods I, IP

3D Features calculation – Gray level distribution, Co-Occurrence matrix measures, etc.

Methods II, AI

Develop a Support Vector Machine implementation to do the texture discrimination.

Validation

Against the current 2D AMFM, and the 3D AMFM that I’m developing for my thesis.

Image Data

Segmented airway trees from multi-slice CT, at least 20 TLC/FRC sets

Goals

Use TLC info for guidance to identify more FRC airways and their branch points.

Methods I, IP

Airway tree segment growing in the FRC tree from higher-level-knowledge identified start-points.

Methods II, AI

Use TLC tree for guidance and identification of additional FRC tree growing start-points, likely after TLC/FRC volume registration.

Validation

Perform manual determination of tree continuation in FRC and compare with computer-results of tree add-ons in FRC volumes.

Image Data

Archived, pathology-proven BMRI data from 20 patients. For each patient, four different image sequences were acquired: T1-weighted spin echo, fat-suppressed T2-weighted fast spin echo and dynamic contrast-enhanced 3D rf-spoiled gradient echo.

Goals

Identification of suspicious masses for subsequent classification.

Methods I, IP

Segmentation will take advantage of the intensity enhancement in breast masses when the post-contrast image is compared to the pre-contrast image. An adaptive region threshold algorithm will be applied to the difference image computed by subtracting the pre-contrast and post-contrast image. Additional information from the T1 and T2 weighted sequences will be included to the threshold algorithm.

Methods II, AI

After region segmentation, region classification is applied to distinguish benign masses from malignant masses. Region shape and texture features, such as mass area, perimeter, compactness, spiculation, average boundary roughness, pixel mean, pixel standard deviation, contrast enhancement, T2 signal, and similar measurements will be used to characterize the mass. An artificial neural network (ANN) will be used to classify masses based on these features.

Validation

Against the pathology-determined independent standard.

Image Data

Greyscale image sequences from the LSDCAS project.

Goals

Develop a methodology for the segmentation of mitotic cells and classification of mitosis outcomes.

Methods I, IP

Watershed segmentation on localized distance images of mitotic cells. Possibly using 3D watershed or motion estimation to match cells between frames.

Methods II, AI

Hidden Markov Models used to classify outcomes of mitosis, and assist in region merging of watershed segmentation.

Validation

LSDCAS experiment files consist of multiple fields with several hundred frames. Will train on half of the fields in an experiment and test on the remaining fields. Outcome will be compared to a manual analysis of the experiment.

Image Data

Micro CT or initially CT pulmonary images in mice - at least 10 mice

Goals

Segment lungs, airway and vascular trees - with usage of programs available for this purpose in humans

Methods I, IP

Methods II, AI

Use the knowledge of general murine anatomy to assist in the segmentation and labeling of lungs (left/right), airway trees, vascular trees (not separated arterial/venous), cardiac region, and lower thoracic cavity.

Validation

Compared to manual tracings, ability to identify individual organs reasonably, as well as quantitative validation of lung surface = positioning errors

Image Data

Sheep pulmonary CT - at least 10 sheep, scanned about 5 times each

Goals

Identify regions of emphysema in 3D, staging of disease ==> generate 3D maps of disease + stage

Methods I, IP

Identify airway and vascular trees, and exclude them from texture analysis of parenchyma

Methods II, AI

Validation

Against manually defined independent standard, possibly using leave-one-out approach

Image Data

Microscopy images of living cells

Goals

Develop a methodology for the segmentation of the all cells in the miscroscopic filed of view

Methods I, IP

Methods based on thresholding, watershed, and mathematical morphology.

Methods II, AI

Use knowledge of cell shape, appearance properties, and the information about context along the time axis

Validation

Comparison with manual segmentation results.

Image Data

Human pulmonary CT images - at least 10 data sets

Goals

Determine whether airway and/or vascular trees cross the boundaries of the lung lobe, use this information for more detailed detection of airway and vascular trees in higher generations

Methods I, IP

Using existing techniques, identify lungs, lobes, airway and vascular trees. Design a method for detecting whether airways or vessels cross regions between lobes. Identify regions of lobes that have insufficient = unreasonably low density of airways and vessels.

Methods II, AI

Resolve conflict of airways/vessels crossing from lobe to lobe. Design an approach of searching for additional airways and vessels in low-density regions.

Validation

Assess correctness of airway/vessel segment removal and addition by comparison to judgment of an expert observer. Likely start working in simple phantom data.

Image Data

True color bronchoscopy images of pulmonary airways. Chest CT scans

Goals

Methods I, IP

Convert RGB color images to 'HSI'. Threshold 'I' to provide zero level set for SFS.

Methods II, AI

Validation

Develop phantom images for analysis of SFS algorithm. Use half of patient data for training of disease classification and validate with remaining data.

Image Data

MR images of cardiac structures 2D possibly 3D - at least 10 data sets with 16 phases each.

Goals

Use segmentation of heart to improve image quality and simultaneously minimize reconstruction time of reconstruction from k-space data

Methods I, IP

Sequence of MR scans, likely 16/s or so. After acquiring 1st phase, reconstruct with no or minimal a priori knowledge. Segment heart or other moving structures.

Methods II, AI

Identify the areas of the scan that may be sampled less densely for the remaining scans to maximize image quality while minimizing scanning time.

Validation

Design a criterion for image quality. Design a criterion for scanning speed. For varying parameters, demonstrate how to optimize scanning time/quality - using some kind of ROC analysis.

Image Data

Goals

Methods I, IP

Methods II, AI

Validation

Image Data

Goals

Methods I, IP

Methods II, AI

Validation

Image Data

Goals

Methods I, IP

Methods II, AI

Validation