A Fully Automatic Global Gradient Measure Based 3D Region Growing Solid Tumour Segmentation Method (3D-GGM-RG) for Low Contrast and Low Count Positron Emission Tomography

Background: Due to low signal-to-noise ratio (SNR) and poor spatial resolution of PET camera along with the finite image sampling constraint, segmentation of solid lesion on PET images is a challenging task. The problems of PET image segmentation even becomes more challenging as the size, contrast and noise of the lesions are subject to vary within and between patient and radiotracers as well as before and after treatment. Most of the methods perform well when the contrast is high and noise is low. However, they are not robust across different statistical fluctuations particularly when the tumour contrast is low. Methods: This paper proposes a novel global gradient measure based 3D region growing (3D-GGM-RG) PET image segmentation method that does not depend on the image statistics. The method is fully automatic and can be easily implemented. The performance of the algorithm is compared with the most widely investigated 40% fixed threshold based method using quantitative measures such as segmented volume, dice similarity coefficient (DSC) and percentage classification error (% CE) in comparison to the true volume with the torso NEMA phantom that contains six different sizes of spheres. The phantom data were acquired with a Siemens TrueV PET-CT scanner at a contrast level of 2:1 between the spheres and background. Results: At contrast level 2:1, 40T method significantly overestimates the volume (≈4.5 times compared to the true volume) and the overestimation is very much dependent on the noise level. On the other hand, the segmented volumes match closely with the true volumes for the 3D-GGM-RG across different noise levels. Average DSC and % CE for the 40T method is 0.32 (ranging from 0.11 to 0.56) and 700% with DSC being dependent on the noise levels. Conversely, average DSC and % CE are 0.70 (ranging from 0.60 to 0.81) and 64% respectively for the 3D-GGM-RG method demonstrating less noise dependency. Conclusion: The performance of the proposed 3D-GGM-RG method is significantly improved and can provide accurate segmentation results across different sizes and noise levels. Thus, the proposed method can be applied where the size of the lesion, contrast and noise can change due to change in radiotracer uptake as responses to treatment. The method can also be used without modification for different radiotracers where variable uptake are expected due to differences in molecular pathways of the tracers.