Segmentation of heterogeneous or small FDG PET positive tissue based on a 3D-locally adaptive random walk algorithm

• We propose to use the distance between adjacent nodes in place of a constant β parameter to take into account the different distances between adjacent nodes in 26 connectivity.
• To reduce the partial volume effect in PET imaging, we propose to strengthen the grouping of voxels having similar intensity by adding the likelihood of probability to each class (tumor and non-tumor).
• The accuracy in the small and heteregenous tumor segmentation is improved using our improvements.

A segmentation algorithm based on the random walk (RW) method, called 3D-LARW, has been developed to delineate small tumors or tumors with a heterogeneous distribution of FDG on PET images. Based on the original algorithm of RW [1], we propose an improved approach using new parameters depending on the Euclidean distance between two adjacent voxels instead of a fixed one and integrating probability densities of labels into the system of linear equations used in the RW. These improvements were evaluated and compared with the original RW method, a thresholding with a fixed value (40% of the maximum in the lesion), an adaptive thresholding algorithm on uniform spheres filled with FDG and FLAB method, on simulated heterogeneous spheres and on clinical data (14 patients). On these three different data, 3D-LARW has shown better segmentation results than the original RW algorithm and the three other methods. As expected, these improvements are more pronounced for the segmentation of small or tumors having heterogeneous FDG uptake.