Abstract
1213
Background: Segmentation of malignant lesions on FDG PET-CT is one of the most expected tasks for deep learning. To develop an efficient deep-learning-based system, a massive database of training data, usually defined by radiologists, is of great importance. Compared to CT and MRI, definition of tumor boundary on PET is easier as various methods of automated segmentation, such as fixed/adaptive threshold and gradient methods, can work. However, time-consuming manual operation that usually has low reproducibility for tumors adjacent to physiological or inflammatory uptake, resulting in degraded reproducibility. Here, we propose a new method that requires minimal human interaction to delineate tumor uptakes, and to test its inter-operator reproducibility.
Methods: In this retrospective study, we focused on segmenting the primary tumor of head-and-neck cancer and gynecological cancer as they frequently co-exist with physiological accumulation. The study population consisted of 23 patients (13 with head-and-neck cancer and 10 with gynecological cancer) of whom FDG PET-CT showed a metabolically active primary tumor. The FDG PET-CT were analyzed using the proposed method as shown in Figure 1. Briefly, all the voxels satisfying SUV ≥ 2.5 were first extracted. Next, the operator specified a point in the 3-D space that was clearly inside the primary tumor. In case the primary tumor was located closely to other components (i.e., metastatic lesion, inflammatory, or physiological uptake), the operator specified a point that was clearly inside one of these other components. The minimum threshold was exhaustively searched to separate the two points. The voxels below the threshold were labeled using a steepest uphill algorithm. In case that an optimum threshold was not found, a nearest neighbor algorithm was used to solve the label collision. The process was repeated until the operator was satisfied with the segmentation result. To evaluate inter-operator reproducibility, the method was used independently by 2 experienced nuclear medicine physicians. Dice similarity coefficient (DSC) was employed to compare the 2 volumes of interest (VOI) of each tumor, calculated by DSC = 2 |VOIA ∩ VOIB| / (|VOIA| + |VOIB|), where VOIA and VOIB were VOI defined by thetwo physicians, respectively. |X| indicates the number of voxels of region X. DSC ranges 0 to 1, and a higher value represents a higher similarity of the two VOIs (i.e., better reproducibility).
Results: The method worked successfully for all the patient images. DSC between VOIA and VOIB (by 2 physicians, respectively) was 0.98 ± 0.03 (mean ± SD), ranging from 0.91 to 1.00, indicating high similarity of the VOIs. Of 23 patients, DSC was 1.0 (i.e., perfect match) in 14 patients (6/13 head-and-neck cancer and 8/10 gynecological cancer). The corresponding metabolic tumor volume (MTV) was 96 ± 127 mL vs. 97 ± 126 mL (VOIA vs. VOIB) with Pearson’s correlation coefficient R being 0.998. Similarly, total lesion glycolysis (TLG) for VOIA and VOIB were 580 ± 720 mL and 584 ± 723 mL, respectively (R = 0.999).
Conclusions: We proposed a new semi-automated method to segment tumors on FDG PET-CT, and demonstrated that the method successfully separated the tumors from physiological or inflammatory uptakes with high inter-operator reproducibility. This method, requiring minimal human interaction, will not only help prepare training data for FDG PET-CT, but will also contribute to conventional analysis, such as measurements of MTV and TLG, and radiomic analysis of tumors.