Comparison Of Segmentation Techniques For NonSmall Cell Lung Cancer (NSCLC) In PET Images

Abstract

Purpose-Tumor segmentation is an important step in cancer disease Computer Aided Diagnosis
(CAD) tool that affects the accuracy of the overall system. In specific, lung tumor segmentation in PET images
is crucial to determine the level of malignancy and the volume of interest. As manual segmentation is time
consuming, laborious and is vulnerable to expert variability, the paper aims to compare three established
segmentation methods namely region growing, fuzzy c-mean and level set method for Non-Small Cell Lung
Cancer (NSCLC) segmentation in PET image. Experimentally, the three methods were tested on 10 PET
images by initially defined the bounding box that contains the tumor. To evaluate the segmentation’s accuracy,
the segmented structures were examined with respect to the manual segmentation as ground truths in terms of
Accuracy, Dice coefficient, Jaccard coefficient, False Positive Ratio (FPR), False Negative Ratio (FNR), True
Positive Ratio (TPR) and True Negative Ratio (TNR). The results show that fuzzy c-means segmentation
method produces the best segmentation results compared to the other two with an accuracy of 89.65% as well
as Jaccard and Dice coefficients 0.82855 and 0.70715 respectively.