Segmentation of lung lesion volume by adaptive positron emission tomography image thresholding.
Academic Article
Overview
abstract
BACKGROUND: It is common protocol in radionuclide therapies to administer a tracer dose of a radiopharmaceutical, determine its lesion uptake and biodistribution by gamma imaging, and then use this information to determine the most effective therapeutic dose. This treatment planning approach can be used to quantitate accurately the activity and volume of lesions and organs with positron emission tomography (PET). In this article, the authors focus on the specification of appropriate volumes of interest (VoI) using PET in association with computed tomography (CT). METHODS: The authors have developed an automatic image segmentation schema to determine the VoI of metastases to the lung from PET images, under conditions of variable background activity. An elliptical Jaszczak phantom containing a set of spheres with volumes ranging from 0.4 to 5.5 mL was filled with F-18 activity (2-3 microCi/mL) corresponding to activities clinically observed in lung lesions. Images were acquired with a cold background and then with variable source-to-background (S/B) ratios of: 7.4, 5.5, 3.1, and 2.8. Lesion VoI analysis was performed on 10 patients with 17 primary or metastatic lung lesions, applying the optimum threshold values derived from the phantom experiments. Initial volume estimates for lung lesions were determined from CT images. Approximate S/B ratios were obtained for the corresponding lesions on F-18-fluoro-2-deoxy-D-glucose (18FDG)-PET images. From the CT estimate of the lesion size and the PET estimate of the S/B ratio, the appropriate optimum threshold could be chosen. The threshold was applied to the PET images to obtain lesion activity and a final estimate of the lesion volume. RESULTS: Phantom data analysis showed that image segmentation converged to a fixed threshold value (from 36% to 44%) for sphere volumes larger than 4 mL, with the exact value depending on the S/B ratios. For patients, the use of optimum threshold schema demonstrated a good correlation (r = 0.999) between the initial volume from CT and the final volume derived from the 18FDG-PET scan (P < 0.02). The mean difference for those volumes was 8.4%. CONCLUSIONS: The adaptive thresholding method applied to PET scans enables the definition of tumor VoI, which hopefully leads to accurate tumor dosimetry. This method can also be applied to small lesions (<4 mL). It should enable physicians to track objectively changes in disease status that could otherwise be obscured by the uncertainties in the region-of-interest drawing, even when the scans are delineated by the same physician.
