Lung Ventilation

lungventilation fig1.jpg

Project Info

Title Classification of Lung Ventilation in Hyperpolarized He-3 MR Images
Goal Design and build a fully automated program that classifies regions of the lung as ventilated, hypo-ventilated, and unventilated


To classify areas of the lung as ventilated, hypo-ventilated, and unventialted.


A new, safe imaging technique has recently been developed that allows doctors to view the ventilation of lungs with a high degree of detail. The techniquie requires a patient breathe in Hyperpolarized He-3 (HH3) gas, which is then imaged on a magnetic resonance (MR) scanner. The scanner is adjusted so that it only detects the presence and quantity of HH3 gas. The greater the quantity of HH3 detected, the brighter the pixel in the image. Doctors can analyze these images to determine the percentages of the lung that are ventilated, hypoventilated, and unventilated.

One drawback however is that the images are time consuming and difficult to analyze. Another problem is that doctors view the images and exhibit biases in analyzing data sets. Inconsistant multiple readings can lead to analysis errors. An automated computer algorithm that analyzes these images would greatly simplify the process and remove human error.

Previous attempts at classifying the images have met limited success. All MR images are corrupted by intensity inhomogeneities, a gradual variation in intensities across an image due to the scanning process and not to the object being imaged. Figure 1 shows examples of the artifacts mentioned. The right lung appears brighter than the left, and there is a fading in brightness at the periphery. This fading is due to the imaging process -- the function of the MRI -- and not to the lung being imaged.

lungventilation fig1.jpg
Fig 1. Example Artifacts.

Standard classification algorithms do not adjust for the intensity inhomogeneities, and therefore fail to properly classify HH3 images. Classification algorithms tested include K-Means, Fuzzy C-Means, Hierarchical K-Means, and Adaptive Fuzzy C-Means. Adaptive Fuzzy C-Means was proposed by Pham and Prince of Johns Hopkins University and showed success in classifying MR images of the brain. However, AFCM has not yet been able to classify HH3 images.

HH3 images cannot distinguish between unventilated lung and other organs. Therefore, HH3 images cannot be used alone to determine the area of the lung that is not ventilated. Figure 2 shows a standard proton MR image. Such images are used to determine the potential lung area. Snakes are capable of segmenting the lungs from the proton image. Preprocessing is requist in order to remove noise. Without any preprocessing of the image, the snake will fail to grow and capture the entire lung because of noise in the image.

lungventilation fig2.jpg
Fig 2. Standard Proton MR Image.

The inclusion filter, developed by current Team VIVA member Nilijan Ray, is run over the image to remove noise. The inclusion filter is a levelset method that only retains connected components which contain or surround pixels of interest.

The objective of this research is to develop an algorithm that segments the lung from the standard proton MR image, classifies the HH3 image according to degrees of ventilation, and calculates the percentage of the lung that is unventilated, hypo-ventilated, and ventilated. The algorithm must be robust to both noise and intensity inhomogeneities.

A GUI was developed to test different classification and segmentation algorithms. Figure 3 is a screen shot of the application. So far, use of the inclusion filter followed by a GGVF snake has produced good results in segmenting out the lungs from the proton image. However, clinical worthy classificaiton techniques have yet to be developed that hold the ability to accurately classify HH3 images

lungventilation fig3.jpg
Fig 3. Classification and Segmentation.


Topic revision: r3 - 30 May 2014, AndreaVaccari
©2018 University of Virginia. Privacy Statement
Virginia Image and Video Analysis - School of Engineering and Applied Science - University of Virginia
P.O. Box 400743 - Charlottesville, VA - 22904 - E-Mail