Contrast-oriented seed based automatic segmentation algorithm: Minimizing effect of lesion heterogeneity on algorithm response

The primary aim of this work was to evaluate the impact of different parameters on the performance of a contrast oriented (COA) seed-based segmentation algorithm for tumor delineation on PET image. In addition, an extended version of the algorithm (2I-A), accounting for lesion heterogeneity, was proposed and preliminarily evaluated. 80 4D-PET images, corresponding to 8 lung cancer patients (10 time-bins/patient), were employed for the analysis. Consensus of three visual PET contours by experts was established as the ground truth surrogate (GTS). The accuracy of the segmented volume by COA was evaluated with respect to GTS using Dice Similarity Coefficient (DSC
_A
). Variability among the experts was also assessed (DSC
_E
). Increasing DSC
_A
/DSC
_E
was employed as an indicator of increasing segmentation accuracy. Dependence of DSC
_A
/DSC
_E
with respect to different parameters was investigated. DSC
_A
/DSC
_E
showed linear correlation with respect to lesion activity concentration A
_GTS
(r=-0.827, slope of m=-0.05) and its homogeneity 1/COV
_GTS
(r=0.750, m=0.22). Good volume segmentation accuracy was obtained on average for COA (DSC
_A
=0.71±0.08 vs DSC
_E
=0.78±0.07), but was further improved for 5 heterogeneous (COV>0.3) lesions by use of 2I-A: 2I-A algorithm diminished volume difference from 37±16 to 16±8% and increased DSC from 0.69±0.08 to 0.82±0.05. From the results, lesion heterogeneity could be identified as the parameter with strongest effect on COA-algorithm accuracy. Although COA showed good accuracy, implementing a second algorithm iteration (2I-A) is recommended because 2I-A further improves volume segmentation accuracy in lesions with high uptake heterogeneity.