Brain tumor segmentation using morphological operation and threshold in PET images

Background and aims: This study aimed to propose a model based on a mathematical operationon each slice of PET image to segment the tumors in three-dimensional (۳D) images of patients’brainsMethod: Our experiments were done on the relevant data obtained from the public dataset of theHECKTOR challenge. In our study, only PET images with careful reference to the CT imageswere used. Hence, PET images of ۴۰۸ patients with brain tumors were used to test the model. Weused Non-Local Means (NLM) noise reduction algorithm to minimize the loss of the essential imageinformation and remove its noise. After pre-processing, morphological operations measuredthe similarity of the intensity and edge information in the image after setting areas with the samesized label located. Then, the images were filtered in three steps based on the location field filters,including Gaussian, median, and mean filters. The Dice score and accuracy are used to comparethe manual and predicted segmentation.Results: The Dice score is applied to calculate overlap among segmented outcomes of the proposedmodel with ground truth annotations. Our proposed method achieved an average Dice scoreof ۸۱.۴۷ ± ۳.۱۵ and an accuracy of ۹۴.۰۳ ± ۴.۴۴.Conclusion: The presented algorithm makes it feasible to produce a patient-specific segmentationof the tumor region without manual interaction. In summary, this model may be highly effectivefor segmenting other organs from small amounts of annotated medical images.

کلیدواژه ها:

brain tumor ، segmentation ، morphological operation ، PET images