Feasibility of a semi-automated contrast-oriented algorithm for tumor segmentation in retrospectively gated PET images: phantom and clinical validation

Abstract

PET/CT plays an important role in radiotherapy planning for lung tumors. Several segmentation algorithms have been proposed for PET tumor segmentation. However, most of them do not take into account respiratory motion and are not well validated. The aim of this work was to evaluate a semi-automated contrast-oriented algorithm (COA) for PET tumor segmentation adapted to retrospectively gated (4D) images. The evaluation involved a wide set of 4D-PET/CT acquisitions of dynamic experimental phantoms and lung cancer patients. In addition, segmentation accuracy of 4D-COA was compared with four other state-of-the-art algorithms. In phantom evaluation, the physical properties of the objects defined the gold standard. In clinical evaluation, the ground truth was estimated by the STAPLE (Simultaneous Truth and Performance Level Estimation) consensus of three manual PET contours by experts. Algorithm evaluation with phantoms resulted in: (i) no statistically significant diameter differences for different targets and movements (Δφ = 0.3 ± 1.6 mm); (ii) reproducibility for heterogeneous and irregular targets independent of user initial interaction and (iii) good segmentation agreement for irregular targets compared to manual CT delineation in terms of Dice Similarity Coefficient (DSC = 0.66 ± 0.04), Positive Predictive Value (PPV = 0.81 ± 0.06) and Sensitivity (Sen. = 0.49 ± 0.05). In clinical evaluation, the segmented volume was in reasonable agreement with the consensus volume (difference in volume (%Vol) = 40 ± 30, DSC = 0.71 ± 0.07 and PPV = 0.90 ± 0.13). High accuracy in target tracking position (ΔME) was obtained for experimental and clinical data (ΔME(exp) = 0 ± 3 mm; ΔME(clin) 0.3 ± 1.4 mm). In the comparison with other lung segmentation methods, 4D-COA has shown the highest volume accuracy in both experimental and clinical data. In conclusion, the accuracy in volume delineation, position tracking and its robustness on highly irregular target movements, make this algorithm a useful tool for 4D-PET based volume definition for radiotherapy planning of lung cancer and may help to improve the reproducibility in PET quantification for therapy response assessment and prognosis.

FDG-PET Radiomics for Response Monitoring in Non-Small-Cell Lung Cancer Treated with Radiation Therapy

“…When comparing the same cohort but with different segmentation approaches, manual segmentation resulted in a larger number of IF satisfying the two criteria (65 normal-distributed and 83 comparable) than for COA (61 normal-distributed and 69 comparable). The decreasing number of robust IF for COA could be justified by the fact that, in comparison with manual segmentation, automatic segmentations like COA are more sensitive to image noise, heterogeneity and signal blurring due to the lesion motion [20]. When comparing two different patient cohorts, more IF were robust between 3D and 4D in cohort 3 (131 IF) than for the same segmentation in cohort 1 (83 IF).…”

“…Two different methods were applied to delineate the primary tumor lesion: (i) a manual contour by consensus of two radiation oncologists and (ii) the semi-automatic segmentation method Contrast-oriented-algorithm (COA) approved by a radiation oncologist. COA was previously validated with heterogeneous experimental phantoms and lung cancer patients [20].…”

“…Retrospectively respiratory gated 4D PET/CT has been proposed to minimize respiratory motion degradation in PET/CT systems [19]. Our group previously investigated the IF variability for 3D and 4D PET imaging protocols with experimental heterogeneous phantoms [20] and lung cancer patients [21]. Results showed that although IF variability across the breathing phases depends on the considered IF, all IF followed a normal distribution for most lesions.…”

Simultaneous Truth and Performance Level Estimation Method for Evaluation of Target Contouring in Radiosurgery

“…Thus, several sets of 6 to 11 contours randomly selected from the total of 12 available for each case were considered, thus investigating the impact of removing between 1 and 6 randomly chosen contours. In addition to the true volume calculated using the maximum number of input volumes of 12 for each case, the true volume was calculated 250 times with varying number of segmentations used as input (6)(7)(8)(9)(10)(11). Minimum, maximum and mean true volumes were recorded for the various simulations after the random removal of a variable number of input segmentations.…”

Experimental phantom evaluation to identify robust positron emission tomography (PET) radiomic features

“…Furthermore, automatic algorithms for lesion segmentation on PET images are preferable in order to minimize the inter- and intra-user variability and their invested time. However, the simplicity and velocity of PET segmentation algorithms are usually in expenses of a loss in segmentation accuracy, especially for heterogeneous lesions [ 6 , 14 ]. Any loss in the segmentation accuracy implied by simple automatic PET segmentation approaches could translate into PET radiomic features not being reliable enough for an accurate quantification of lesion heterogeneity.…”