A gradient-based method for segmenting FDG-PET images: methodology and validation

Abstract

Purpose A new gradient-based method for segmenting FDG-PET images is described and validated. Methods The proposed method relies on the watershed transform and hierarchical cluster analysis. To allow a better estimation of the gradient intensity, iteratively reconstructed images were first denoised and deblurred with an edge-preserving filter and a constrained iterative deconvolution algorithm. Validation was first performed on computer-generated 3D phantoms containing spheres, then on a real cylindrical Lucite phantom containing spheres of different volumes ranging from 2.1 to 92.9 ml. Moreover, laryngeal tumours from seven patients were segmented on PET images acquired before laryngectomy by the gradient-based method and the thresholding method based on the source-to-background ratio developed by Daisne (Radiother Oncol 2003;69:247–50). For the spheres, the calculated volumes and radii were compared with the known values; for laryngeal tumours, the volumes were compared with the macroscopic specimens. Volume mismatches were also analysed. Results On computer-generated phantoms, the deconvolution algorithm decreased the mis-estimate of volumes and radii. For the Lucite phantom, the gradient-based method led to a slight underestimation of sphere volumes (by 10–20%), corresponding to negligible radius differences (0.5–1.1 mm); for laryngeal tumours, the segmented volumes by the gradient-based method agreed with those delineated on the macroscopic specimens, whereas the threshold-based method overestimated the true volume by 68% (p = 0.014). Lastly, macroscopic laryngeal specimens were totally encompassed by neither the threshold-based nor the gradient-based volumes. Conclusion The gradient-based segmentation method applied on denoised and deblurred images proved to be more accurate than the source-to-background ratio method.