An Investigation of Iterative Reconstruction Strategies for Lung Lesion Detection in SPECT H.C. Gifford 1 , X.M. Zheng 2 , G. Boening 1 , P.P. Bruyant 1 , and M.A. King 1 1 University of Massachusetts Medical School, Worcester, MA, USA 2 Charles Sturt University, Wagga Wagga, NSW, Australia Abstract— ROC and localization ROC (LROC) studies assessed whether modeling of the acquisition physics during the recon- struction process could improve detection of SLN in SPECT images. The radiotracer used was Tc-99m-labeled Neotect, and studies were run both with hybrid images and with phantom images. Iterative reconstruction strategies were defined with various combinations of attenuation correction (AC), scatter correction (SC), and resolution correction (RC). Results from both human and model observers indicate that AC degrades lung- lesion detection relative to applying RC alone. Subsequent studies with the model observer suggest that the relative effectiveness of the different iterative strategies will depend on observer knowledge of the mean normal background. I. I NTRODUCTION Solitary lung nodules (SLN) show up frequently in radio- graphic scans, and while the majority of them are benign, they can represent early stages of lung cancer. In such cases, the chances of a cure are greatly improved by timely diagnosis. Our goal was to determine the extent to which detction of SLN could be improved by proper modeling of the acquisition physics during the reconstruction process. The imaging pro- tocol we considered was for Tc-99m-labeled NeoTect SPECT imaging, which has been reported to offer performance com- parable to FDG-PET for evaluating SLN [1]. For our evaluations, ROC and localization ROC (LROC) observer studies were performed with human and model ob- servers. As NeoTect has mainly been used to evaluate certain nodules in patients who have tested positive with other imaging tests, we conducted ROC studies with a “location-known- exactly” (LKE) detection task, in which the prospective lesion location in a given image is prior information. A localization component was added to the detection task by requiring the observers to search for the lesions within the entire region of the lungs. Studies based on this task were run using LROC methodology [2]. A channelized nonprewhitening (CNPW) multiclass model observer was capable of performing this localization-detection task. Our primary studies were done with hybrid images. Such images present simulated abnormalities within clinically nor- mal patient backgrounds, offering greater realism compared to This work was supported by the National Institute for Biomedical Imaging and Bioengineering under grant number R01-EB02798. The contents are solely the responsibility of the authors and do not necessarily represent the official views of the NIBIB. using purely mathematical phantoms while still allowing the experimenter to know the truth of the cases involved. Zheng et al. [3] introduced our use of NeoTect hybrid images in pilot LROC studies aimed at optimizing the parameters for filtered backprojection (FBP) and rescaled-block-iterative (RBI) recon- struction strategies. The RBI strategy in [3] applied iterative attenuation correction (AC), resolution correction (RC), and scatter correction (SC). As part of this current work, we tested the same strategy but also enlisted RBI strategies featuring subsets of those corrections. Additional observer studies were carried out on recon- structed images of a mathematical phantom. These phantom studies allowed us to examine several issues that arose from the hybrid-image studies but which could not be performed with those images. II. METHODS A. Data Acquisitions 1) Clinical Data: NeoTect scans from nine patients were used for the hybrid images. The emission and transmission data for a patient were acquired simultaneously on a Philips Prism 2000XP with low-energy, ultra-high-resolution collimators and a Gd-153 scanning-line source. Emission counts were collected with a 20% photopeak window at 140.5 kev and scatter win- dows at 123 kev (4.9%) and 159 kev (3.8%). Windows for the transmission data were set at 99.0 kev (25%; photopeak), 83.0 kev (7%; scatter) and 116.5 (5.2%; scatter). The protocol called for a 24-minute scan with 120 128×128 projections over 360 degrees and the total numbers of emission photopeak counts per patient ranged from 9.5 M to 23.0 M. Transmission maps of dimensions 128 3 (0.467-cm voxel width) were reconstructed with the Philips iterative software. Downscatter and uniformity corrections of the transmission data were implemented. Each patient volume was used to create 18 abnormal cases with one lesion each. These lesions were spheres of 1-cm diameter and they were randomly placed within the lung areas as determined from a segmentation of the transmission maps. Projections of the lesions were generated with the SIMIND [4] Monte Carlo program, using the attenuation weights supplied by the clinical transmission maps. The relative lesion-to- background contrasts were scaled to produce average areas under the LROC curve of about 0.75 for the human observers. 0-7803-8701-5/04/$20.00 (C) 2004 IEEE