Abstract:
Tumor segmentation accuracy greatly affects the effectiveness of radiotherapy procedures. Maximizing the segmentation accuracy has high medical significance in order to deliver the highest radiation dose to the target volume while protecting the healthy tissues. This dissertation aims to present an optimized method to minimize errors in the automated segmentation of tumors in PET images. Blind deconvolution wasimplementedinaregionofinterestencompassingthetumorwithaniterationnumber determined from Contrast-to-Noise Ratios. The images were resampled. Several automaticsegmentationalgorithmsweretestedonthreedatasets: phantom, simulated geometric lesions inserted in real images, and simulated clinical images with real heterogeneous tumors for which ground truth was known. The volumes of the tumors were 0.49-26.34cc, 0.64-1.52cc, and40.38-203.84ccrespectivelyforthethreedatasets. The widely available software tools MATLAB, MIPAV, and ITK-SNAP were used. With theuseoftheactivecontourwithclassificationtechnique,themeanerrorswerereduced from 95.85% to 3.37%, from 815.63% to 17.45%, and from 32.61% to 6.80% for all the lesionsofthephantomdataset,thesimulateddataset,andthelargelesionsoftheclinicalPETdatasetrespectively. Thecomputationaltimewasreducedbyafactorofmore than 10 by the use of region-of-interest-based deconvolution. Contrast-to-Noise Ratio and Region-of-Interest based deconvolution have the potential to improve delineation accuracyfordifferentsizesofhomogeneousandheterogeneoustumors. Improvementis veryimportantforsmallertumors. Thealgorithmmayprovidereducedcomputational time with respect to full deconvolution and can be implemented using widely available software tools.|Keywords : PET, medical image segmentation, image restoration, blind deconvolution, radiotherapy planning.