RT Journal Article SR Electronic T1 An estimation-based segmentation method to delineate tumors in PET images. JF Journal of Nuclear Medicine JO J Nucl Med FD Society of Nuclear Medicine SP 447 OP 447 VO 61 IS supplement 1 A1 Ziping Liu A1 Richard Laforest A1 Hae Sol Moon A1 Joyce Mhlanga A1 Tyler Fraum A1 Malak Itani A1 Aaron Mintz A1 Farrokh Dehdashti A1 Barry Siegel A1 Abhinav Jha YR 2020 UL http://jnm.snmjournals.org/content/61/supplement_1/447.abstract AB 447Objectives: Accurate tumor delineation in oncological PET images is important for PET-based radiotherapy planning and quantification of radiomic and volumetric features [1]. However, PET segmentation is challenging, primarily due to partial-volume effects (PVEs) arising from low system resolution and tissue-fraction effects (TFEs) [2]. Conventional PET segmentation methods yield limited performance due to the high noise and low resolution in PET images. Further, these methods do not account for TFEs accurately since they are classification-based, i.e. they exclusively classify each voxel as either tumor or normal tissue. To address these issues, our objective was to develop an estimation-based segmentation method that estimates the tumor-fraction area within each pixel to yield the tumor boundaries. Methods: We designed a Bayesian technique that, by minimizing the binary cross-entropy loss between the estimated and true tumor-fraction area over a population of images, estimates the posterior-mean of the tumor-fraction area within each pixel in the reconstructed image. The technique was implemented using an auto-encoder. Evaluating the method required knowledge of the ground-truth tumor boundaries. For this purpose, a realistic simulation framework was designed that generated high-resolution clinically realistic tumor models using patient-data-derived tumor properties (size, shape, uptake values) and simulated intra-tumor heterogeneity using a stochastic lumpy model. A projection-domain-based lesion-insertion approach (Fig. 1a) [3] was used to simulate PET images containing these tumors (Fig. 1b). Tumor realism in the simulated images was evaluated by five nuclear medicine physicians and one nuclear medicine physicist using a two-alternative-forced-choice (2-AFC) study [3]. The readers were shown 50 pairs of PET images, one with a real and the other with a simulated tumor, and asked to identify the image with the real tumor and provide confidence in their decisions. Following this validation, images generated using the simulation framework were used to quantitatively evaluate the proposed segmentation method. The evaluation metrics included fuzzy Dice and Jaccard Similarity Coefficients (fDSC and fJSC), Area Similarity (AS), and Hausdorff Distance (HD). Results: All trained readers achieved a percent accuracy close to 50% with overall low confidence in the 2-AFC study, providing evidence of the realism of the simulated images (Table 1A). The proposed segmentation method significantly outperformed (p < 0.01) other considered methods, including a U-net-based and a fuzzy segmentation approach, on the basis of all considered metrics (Fig. 2 and Table 1B), e.g., the method yielded the highest fDSC value of 0.80 (95% CI: 0.80, 0.81). The method was also observed to be relatively insensitive to PVEs (Fig. 3). Qualitatively, the method yielded an accurate estimate of the ground-truth tumor boundaries (Fig. 4). Conclusions: The proposed estimation-based segmentation method yielded reliable tumor segmentation as rigorously evaluated using trained-readers-validated highly realistic simulation studies. The results motivate the validation of this method using physical-phantom and clinical studies. Further, the results indicate the potential of this method in defining high-resolution boundaries from low-resolution images.