Image Data

3D images of tibia at the knee joint

Goals

Segmentation of the tibia bone, and possibly other bones present in the image + leg outer surface

Methods I, IP

Mean shift segmentation, graph cut segmentation

Methods II, AI

Consider the leg outline + presence of other bones in the neighborhood to increase segmentation robustness

Validation

12 3D human data sets, manual tracings in at least 15 slices per data set, perform quantitative validation. Compare mean shift/graph cut, and your standard fuzzy connectivity results (if available). Show that Methods I truly benefirt from adding Methods II to the segmentation procedure.

Image Data

Goals

Artery-vein separation in pulmonary CT images

Methods I, IP

Extend current methods for vessel segmentation (Sato, Shikata) to large vessels connecting lungs and heart.

Methods II, AI

Use labeling (identification) of these large vessels as a contributing information for artery - vein separation algorithms.

Validation

Using the cast + manually identified arteries/veins in 3 datasets as independent standard , validate your approach in these 3 datasets + the cast, and demonstrate the performance in at least 15 other data sets based on visual assessment.As part of this assessment, assess whether your arterial/venous trees are free of loops and connections - provide quantitative analysis of these phenomena if present.

Image Data

Goals

Reconstruct missing slice information from adjacent slices in trabecular bone

Methods I, IP

Segment bone primitives, describe primitives, determine their 3D connectivity properties, as well as 2D/3D inter-relationships.

Methods II, AI

Using machine learning approaches, learn the globally optimal relationships between primitives in slice 1 and slice 3 considering knowledge of primitives and their properties in slice 2. Use this learned knowledge to estimate (reconstruct) the primnitives, their properties, and configurations in slice 2 from slices 1 and 3 only. If successful, explore the same approach for slices 1+4, 1+5 etc.

Validation

Quantitatively compare properties of primitives and their inte-linking relationships from the reconstructed slices with those of the original slices. Validate in at least 12 3D human data sets, and in at least 15 slices per data set.

Image Data

Goals

Cartilage segmentation in full 3D

Methods I, IP

Multi-surface graph search for interacting surfaces - simultaneous segmentation of tibia/femur/patella.Explore the use of mean shift and/or graph cut for initial bone segmentation.

Methods II, AI

Incorporate fuzzy connectivity and/or other clustering/mean-shift approaches in the cost function design of the final multi-surface segmentation. Explore uise of fuzzy segmentation that maintains prescribed topology for final cartilage segmentation.

Validation

Apply your approaches to at least 20 OARSI MR datasets. Create an independent standard for at least 10 slices in each of teh 20 data sets. Perform quantitative validation of your approaches, compare their performance using surface positioning errors, as well as the start/end cartilage assessment.

Image Data

Goals

Liver segmentation

Methods I, IP

Build an ASM or AAM for liver segmentation - using 30 liver segmentations - based on existing AAM tools developed by Honghai Zhang for hearts. Apply this model to the original CT data and or to the chunk-based segmentation. The goal is to inmprove the currrently available segmentations of the liver.

Methods II, AI

Develop an approach for conflict resolution (conflict between the existing chunks and the smooth model-based AAM/ASM segmentation).

Validation

Apply the developed approaches to the 30 CT images for which liver segmentation independent standard exists, perform quantitative validation of teh segmentation outcome, demonstrate that your approach is better than any segmentation directly resuilting from the chunk-based detection (considering the possible varieties of liver surfaces for the chunks intersecting the true liver surface only)

Image Data

Goals

4D segmentation of IVUS data - internal/external elastic lamina

Methods I, IP

2-surface graph search in 4D - develop novel cost functions

Methods II, AI

Learning-based cost function design, incorporate feasible motion properties of coronary arteries in the segmentation process

Validation

Validate in comparison with independent standard on at least 40 IVUS 3D datasets for which manual tracings are available. Validate on at least 20 4D datasets.