55:247 Image Analysis and Understanding

Project Listing - Spring 2010


IP = Image processing part

AI = Knowledge-based part


Name: William Monroe

Image Data

Liver CT scans

Goals

Segment livers in 3D, including livers with tumors.

Methods I, IP

Level set segmentation, graph-based segmentation

Methods II, AI

Use a leave-N-out (or leave 10/15/25% out) approach to learn parameters of your model from manual tracings. Explore how to incorporate shape (at least partially) in the segmentation techniques.

Validation

20 liver CT images will be segmented using both approaches, the performance will be compared. Both methods will be validated against available independent standard. Potentially, you can compare performance with different leave-N-out approaches and with different cardinality of the data set.

Name: Michael Anderson

Image Data

Either look for an appropriate database on line or get your own images from your camera.

Goals

Implement a system capable of recognizing road signs from continuous footage.  Ideally the method will be able to properly discern the sign, therefore its meaning, and possibly gauge the distance, which would provide other functions the required information to react accordingly. This "reaction" is not part of this project.

 

Methods I, IP

Try employing invariants for rectangular and circular shapes. Use color information. Use a rough recognition of "street" or "road" to identify where the signs are most likely located.

Methods II, AI

Incorporate conflict resolution for cases recognized with different class assigned earlier than on later frames. Use non-overlapping sets of street signs for training of your methods.

Validation

Starting with static images - say 100 images of 5 sign classes, determine the rate of success and report as a confusion table.

Then, perform a similar analysis for at least 20 video loops consisting of at least 100 frames each.

Name: Shanhui Sun

Image Data

Lung CT images in 3D

Goals

Develop a method for lung lobe segmentation in 3D, using a single model of left/right lung with a model of the pulmonary fissures (one model for left, another model for right lung).

Methods I, IP

Use ASM, AAM, or other modeling approach that works in 3D. You may but do not have to use GPUs.

Methods II, AI

Follow up with an optimal surface detection algorithhm to find accurate surfaces of the lungs. Inside of the lungs, find the approximate positins of the fissuere planes (based on darkness in the smoothed 3D image) and use for defining a regionof interest for fissure finding usiing optimal surface graph search (2nd level graph search)

Validation

Determine segmentation accuracy separately for lungs and lobes. Work with at least 20 3D MDCT lung images for which independent standards are available.

Name: Matineh Shaker

Image Data

3D MDCT pulmonary images

Goals

Segment heart from the lung in ex-vivo mouse X-ray CT images

Methods I, IP

Region growing and active shape model segmentation.

Methods II, AI

Train an ASM model from 6 examples.

Validation

Leave one out training/testing approach, determine performance based on:

- copmpared with manual tracing in 3D

- calculate surface positioning errors and DICE coeffcients

Name: Kaifang Du

Image Data

The image data is 4DCT lung data (we have 4 subjects, 10 phases for each respiratory circle). Let's plan on at least 10 of these if possible. Use sheep datasets as well.

Goals

The goal of this project is to generate a 4D statistical model of respiratory lung motion. My project topic would be "Prediction of Respiratory Motion using Lung Registration".

Methods I, IP

Cross-phase intra-subject registration, generate a statistical model using some meaningful landmarks. Develop a method for automated detection of these landmarks.

Methods II, AI

Derive a predictive model of a breathing cycle - predict location of your landmarks in a subset of phases, learn the model parameters from full sets of phases from a training set.

Validation

Methods will be validated against available independent standard - this standard will have manually identified landmarks - the same landmarks detected automatically. Determine landmark positioning errors for landmark detection and landmark position prediction error for prediction part. Use lave one out approach to maximize use of your datasets.

Name: Panfang Hua

Image Data

20 3D Human Thorax CT Images

Goals

Segment the pathological Lung from the thorax CT images - pathologies include emphysema, fibrosis, nodules, and other parenchymal diseases

Methods I, IP

Layered Graph Cut Algorithm (optimal surfaces, Kang Li), use PASS to provide an initialization. Develop a method for rib recognition in 3D.

Methods II, AI

Incorporate rib location into the cost function, learn properties of the cost function from a training set.

Validation

20 3D human data sets, compared with manual segmentation, with the result of Method1, and with Hu's Method. Show the method are better than Hu's Method.

Name: Xiayu Xu

Image Data

Stereo fundus images

Goals

Segmentation of arteries and veins

Methods I, IP

Use single-camera and stereo fundus images to detect vessels - using a graph-structure representation separate A and V subtrees. Once the subtrees are identified, use edge detection (2D graph search) to determine accurate locations of vessel borders so that the measurements (below) can be determined accurately.

Methods II, AI

Identification/labeling of A and V, measurement of their properties

Validation

50 fundus photography images, validation will be performed in at least 20 images for which manually-defined A/V separation and accurate vessel wall borders will be available.

Name: Minqing Chen

Image Data

30 MVCBCT projection images (mega voltage cone-beam CT)

Goals

Determine the top of hemi-diaphragm position in MVCBCT projections

Methods I, IP

Test deformable model, ASM/AAM or graph-surface based approach to identify the diaphragm in a dynamic image sequence.

Methods II, AI

Use optic flow, shape priors, multi-scale techniques, motion models, etc. in the above approaches - the parameters of associated cost/energy functions shall be learned from training set examples.

Validation

Use independent standards (manual tracings) for validation as well as for the method training - use leave x% out approach to separate training/testing sets.

Name: Cheng Zhang

Image Data

Manually outlined 3D cone beam CT & fan-beam CT

Goals

Uncertainty-weighted surface registration.

Methods I, IP

Image registration - Point Registration, Surface Registration.

Methods II, AI

Build a prostate PCA model and a boundary uncertainty model, the boundary drawn by human observers can be corrected so that
the registration thereafter can be better.

Validation

15 patient datasets, show that your approach is better than an approach without uncertainty in comparison to the independent standard.

Name: Halim Choi

Image Data

Using data base of16 MRI normal brains for which independent standard WM/GM/CSF segmentatin exists

Goals

Develop a segmentation technique based on image registration, perform quantitative analysis of brains not used for development of the segmentation technique.

Methods I, IP

Develop an image segmentation technique based on registration of brains with one specific branin template from 16 MR brain images - perform 16 registrations against 16 different template brains. Select at least 3 different registration approaches.

Methods II, AI

For each of the 16 registrations, you will have 15 transformations, and with corresponding 15 inverse transforms, you will determine errors of segmentation.

Validation

Validate registration-based segmentation by determining errors of WM/GM/CSF percentages in individual brains and for the 16 different templates - report as average +- stdev. Also report the average +- stdev of DICE coefficient for each class = WM/GM/CSF.

Compare results from the at least three different registration approaches used.

Name: Hiep Nguyen

Image Data

3D OCT images

Goals

Develop a Content Based image retrieval system for identification of excellent/good/marginal/bad image quality in OCT images.

Methods I, IP

Finding optimal features for image' tiles carving. (Finding similar images based on choosing appropriate visual features which is not be affected too much by carving technique). Use multi-scale approach ... Breaking the image to small tiles, apply carving technique first, extract visual features.

Methods II, AI

Derive proper features suitable for the quality assessment in OCT from the above-identified features.

Validation

Test individual combination of visual features, choose the best combination which preserve the quality of image. Use CBIR system as a search engine, the input is a carving image, the expected output is the normal images which may related to the input image.

Based on manual assessment of OCT image quality, determine the performance of the new automated technique based on image retrieval approaches.

Name: Qi Song

Image Data

3-D CT images of the bladder and the prostate

Goals

Simultaneous segmentation of the bladder and the prostate

Methods I, IP

Cost function design: Incorporate class-uncertainty theory for arc-weighted graph search. The original theory will be extended to multiple objects in 3-D case, i.e., backgrounds, bladder, prostate, bone.

Methods II, AI

Incorporate shape information into arc-weighted graph search framework using model-based method, e.g., ASM/AAM.

Validation

21 3-D CT images from different patients with prostate cancer will be employed for validation. The segmentation results will be compared with the expert-defined independent standard for quantitative validation. Use in a separate training/testing manner - use leave-x%-out to best use all data.


Last Modified: 2010

[Go Back]