The specific aim of this project is to develop image processing and workflow tools to streamline reporting of lung lesions on CT images. Images of lungs will be used to create “lung atlases” that will automatically identify the lung within the image and establish anatomical coordinate system used for reporting on lung lesions.
A subset of the lung CT image studies will be used as a training set. Lung images will be manually annotated to train a multi-atlas learning-based label fusion method [1,2]. Once trained the software will automatically identify (segment out) the lung within the image and establish an anatomical coordinate system. This will allow the system to automatically suggest the location of a lesion (or Region of Interest) marked by the Radiology thus saving valuable time.
The accuracy of the automated identification will be determined by applying the automated lung detection on datasets that were not part of the training set. The automatically segmented lungs will be compared to manually annotated lungs that were not part of the training data set.
Accuracy of registration to prior lung CT images of the same patient will also be investigated. Successful matching to priors will help radiologists track progression of tumors from priors to the current study.
The study will also look into the speed of lung segmentation and registration to priors. This will determine whether lung segmentation and prior registration must be done as a pre-processing step or whether it is fast enough to perform on-the-fly.

IBM Research
Merge Healthcare an IBM Company