LUNGS SEGMENTATION BY DEVELOPING BINARY MASK

Abstract

CT scans have proved to be an important modality for diagnoses of thoracic diseases such as TB, emphysema and cancer. However, soft tissues and minor pathologies are not clearly identifiable from CT. Computer aided processing of CT data can solve the problem. For computer aided diagnoses of lungs diseases, lungs segmentation is a prerequisite. Moreover, lungs segmentation has its application in visualization, volumetric measurement, shape representation and analysis, image guided surgery and nodule size measurement etc. Most of the lungs segmentation methods are scanner dependent. We proposed and implemented a novel, 3-D fully automated machine independent method for segmenting lungs region from CT slices using hybrid approach. Thresholding, mathematical morphology and pixel connectivity have been utilized to achieve the objective. We approached the problem by generating binary mask which identifies the lungs regions on each CT slice. Then, lungs are segmented using the binary mask. The method of generating mask comprised of five steps. 1) Gray level threshold value has been calculated by iterative method and Otsu's method. Results of both methods are same with difference in decimal places only. Iterative method maximizes within class similarity and Otsu's method maximizes between class variance. Using calculated threshold, thresholded binary image has been generated. 2) All objects are identified on the thresholded image by taking advantage of connectivity of pixels and objects connected with border of the image have been removed. These objects were due to attenuation of X-rays through air around the patient. 3) Gaps on region of interest have been stuffed by morphological filling. 4) Trachea, bronchi's and other areas which could be mistaken for lungs have been removed by exploiting anatomical properties of the lungs. 5) Mask boundaries have been smoothed by morphological closing. Finally, lungs have been segmented by arrays product of final mask generated in first step to fifth step and original CT image. The proposed method has been applied on data set of three complete CT scans (20 slices each) and 25 more slices thus in all on 85 slices taken from two different sources. Results have been compared with manually delineated lungs on CT images by a radiologist. Mean Overlapping Fraction, mean precision, mean Sensitivity/Recall, mean Specificity, mean Accuracy and mean Fmeasure have been calculated and recorded as 0.9900, 0.9962, 0.9966, 0.9977, 0.9992 and 0.9964 respectively.