M5C2  Assessment and Comparison of Image Quality and Methods

Friday, Nov. 6  14:00-16:00  Pacific Salon 1&2

Session Chair:  Dimitris Visvikis, LaTIM, France; Georges El Fakhri, Harvard Medical School and Massachusetts General Hospital, United States

Show/Hide All Abstracts

(14:00) M5C2-1, Do Phantom Harmonization Efforts Translate into Harmonized Patient Images?

J. V. Panetta, J. S. Karp, M. E. Daube-Witherspoon

Department of Radiology, University of Pennsylvania, Philadelphia, PA, USA
Scanners with different performance characteristics and reconstruction protocols can produce images that show large variations in the uptake in small lesions. There have been a number of studies aimed at reducing this variability for clinical trials using phantom data to optimize the acquisition and/or reconstruction protocols. The underlying assumption is that the protocols that harmonize the phantom images will also result in reduced variability in patient images among scanners/sites. The goal of this study was to use patient data with embedded lesions of known uptake to test this assumption. The NEMA IQ phantom with hot spheres (diam.: 10-37 mm) was scanned on TOF PET scanners with different intrinsic performance characteristics and reconstructed with a variety of reconstruction protocols. The phantom data were used to harmonize the contrast recovery coefficient (CRC) in small spheres by choosing the reconstruction protocols to give the smallest variation in CRC. Patient FDG data were also acquired on these scanners and the data stored in list-mode format. List-mode events from spheres measured in air on these systems were embedded into the patient list-mode data to insert 5-6 spheres in the liver and lung regions with a known (10:1) uptake with respect to the average whole-body uptake. The impact of applying the harmonization reconstruction parameters from the phantom on the measured sphere uptake (SUV and CRC) for the patient data was studied. Even when reconstruction protocols that harmonized the phantom images were used for the patient data, the difference in measured SUV between the scanners exceeded the variability seen among sphere locations within an organ. However, CRC values of spheres in the liver were within the inter-sphere variability; those in the lung showed greater difference. This study shows that harmonization of patient data may require more careful attention to fine-tune the phantom guidelines.

(14:15) M5C2-2, Investigation of scan time for solitary pulmonary nodule discrimination

H. Bal, V. Panin, J. Hamill, M. Conti, B. Bendriem, M. Casey

Siemens Medical Solutions - MI, Knoxville, Tn, USA
Current recommendation for using PET imaging in the assessment of solitary pulmonary nodules is limited to sizes greater than 8mm. Our objective was to assess the scan time required in order to improve discrimination of small lesions with realistic FDG uptake using XCAT phantom based simulations. An XCAT phantom of the thorax (with respiratory motion) was created with six spherical lesions (6mm and 8mm diameter with SUV ~ 2 SUV ~ 4). For each dataset, 6 scan times ranging from 2.85 - 42 minutes were created to represent the ungated PET data. The corresponding optimal gated PET data were created by using 35% of total counts that of ungated data. Time-of-flight (TOF) sinograms modelling image acquisition physics were simulated and Poisson noise added to obtain 25 noise realizations for each case. Each dataset was reconstructed with nonTOF PSF (PSF) and TOF PSF (PSFTOF) reconstruction using a matched and free-breathing attenuation map. In addition, joint estimation TOF-MLAA method was used with free-breathing attenuation map. Discrimination of lesions was performed by assigning data with SUV = 2 as "benign" and SUV = 4 as "malignant". Signal-to-noise ratio as well as area under the ROC curve were computed and used as a figure of merit. The maximum joint sensitivity and specificity was also computed from the ROC curve for different methods. It was observed that optimal gating and TOF based reconstruction provided improved performance for all cases. Ungated PET data provided poor performance even when scan time was increased. Optimal gating with TOF reconstruction approached AROC=1 earlier compared to nonTOF based reconstruction. Maximum time was required by nonTOF PSF based reconstruction with optimal gating (42 minutes) and the shortest was with TOF PSF based reconstruction with optimal gating and free-breathing CT (10 minutes). The use of longer scan durations, optimal gating and TOF based reconstructions help in improved discrimination of solitary pulmonary nodules.

(14:30) M5C2-3, High Quality Short Frame Reconstruction in Dynamic PET

W. Zhu1, M. Chen2, Y. Dong1, J. Bao2, H. Li1

1UIH America, Inc, Houston, USA
2MI, Shanghai United Imaging Healthcare, Shanghai, China
Low count PET image reconstruction usually suffers from high image noise and quantitative bias. This is particularly significant for short frames (temporal-ROI, or T-ROI) in dynamic PET studies. We proposed a method to reconstruct short frame by utilizing information from a longer acquisition time containing the short frame. The longer acquisition data excluding the short frame is first sorted to perform a reconstruction for the out-of-T-ROI activity. The out-of-T-ROI image is then forward projected to obtain its contribution in the data space. A second reconstruction is then executed with data from the entire long acquisition and the contribution from the out-of-T-ROI to estimate the activity for the short frame. Results show that the image quality and CRC-noise performance are both improved with our proposed method than the standard reconstruction with counts from the target frame only.

(14:45) M5C2-4, Image Quality Assessment of Multiplexed PET

J. L. Herraiz1,2, S. C. Moore3, M.-A. Park3, J. M. Udias2, J. J. Vaquero4, E. Lage5

1Research Lab. of Electronics, Massachusetts Institute of Technology, Madrid-MIT Consortium, Cambridge, MA, USA
2Grupo de Física Nuclear. Dpto. de Física Atómica, Molecular y Nuclear, Universidad Complutense de Madrid, Madrid, Spain
3Division of Nuclear Medicine. Department of Radiology, Brigham and Women’s Hospital and Harvard Medical School, Boston, MA, USA
4Grupo de Bioingenieria e Ingenieria Aeroespacial, Universidad Carlos III de Madrid, Madrid, Spain
5Instituto de Investigaciones Biomedicas “Alberto Sols”, Universidad Autonoma de Madrid, Madrid, Spain
Positron emission tomography (PET) is one of the most sensitive non-invasive molecular imaging tool, being its sensitivity several orders of magnitude higher than that typically obtained in single photon emission computed tomography (SPECT). However, PET lacks the ability of SPECT to multiplex signals from several tracers, which is very useful in many different studies such as cardiac imaging with 99mTc-Sestamibi and 201Tl. Recently, it has been shown that the use of tracers labelled with positron-gamma emitter radionuclides like (124I, 86Y, 82Rb, 94mTc, 76Br) in combination with tracers labelled with standard positron-emitter radionuclides like (18F, 11C, 13N) enables multiplexed PET (mPET). mPET uses the triple coincidences from the positron-gamma emitters, together with the standard double coincidences to reconstruct separated images of each radionuclide's activity distribution. We obtained encouraging results with mPET in some initial preclinical studies, but a detailed study of the quality and quantification properties of mPET images, and an evaluation of its performance in realistic clinical scenarios was still required. The acquisition software of the Argus preclinical PET/CT scanner was adapted to provide datasets of double and triple coincidences, and then used to acquire image quality phantoms filled with different activity concentrations of 18F and 124I. The quality of the separated images of both radionuclides obtained with mPET was assessed in terms of bias, contrast, resolution and noise. Our results show that with the appropriate corrections for triple randoms and inter-detector scatter, mPET provides images with little bias (<10%), good contrast (95% of that obtained with single tracer acquisitions) and a small relative noise increase (<10%). The simulations of acquisitions in clinical scanners with realistic concentrations of 18F and 124I showed that mPET can be also used in clinical cases for simultaneous PET imaging of several radiotracers.

(15:00) M5C2-5, Improving the Identification of Tumor Sub-Volumes Associated with Residual Disease after (chemo)radiotherapy Through Automatic Segmentation of 18F-FDG PET/CT Images

M.-C. Desseroit1, C. Cheze Le Rest2,1, R. Perdrisot2, R. Guillevin3, D. Visvikis1, M. Hatt1

1LaTIM UMR 1101, INSERM, Brest, France
2DACTIM, nuclear medicine departmen, University Hospital Milétrie, Poitiers, France
3Radiology department, University Hospital Milétrie, Poitiers, France
Purpose: 18F-FDG PET/CT is used for diagnosis, staging, radiotherapy planning and for assessment of response to therapy. It has been recently showed in several cancer types that post-treatment residual PET uptake after (chemo)radiotherapy could be spatially located within the tumor area with the highest uptake in the pre-therapy PET. This strengthens the motivation to deliver higher dose to these specific sub-volumes in order to improve local control and reduce recurrence rates. These studies used arbitrary-set SUV-thresholds, known to have limited accuracy. Our goal was to verify this hypothesis with a more accurate method. Material and methods: 44 patients with head & neck (HNC) or esophageal (EC) tumors with both a pre- and post-treatment 18F-FDG PET/CT scan 3 months after standard (chemo)radiotherapy were analyzed. In 10 HNC and 12 EC patients, a post-therapy residual uptake could be detected. First the two PET images were co-registered. Manual correction was applied in cases of mismatch. The pre-therapy functional tumor volume and its higher uptake sub-volume, as well as the residual uptake in the post-therapy image, were then defined using FLAB and the various combinations of thresholds. The overlap fraction (OF) metric was used to compare the resulting volumes. Results and conclusion: EC cases were easier to register than HNC, in which major anatomical deformations sometimes occurred. In EC, the pre-treatment high uptake sub-volume overlapped well with the post-therapy residual uptake (OF=0.64). On the contrary, OF were <0.5 for half of the HNC cases, which could be explained by organs/tumor deformation. OF were also systematically higher with FLAB than with any combination of SUV thresholds. Our results strengthen the hypothesis that high uptake sub-volumes could benefit from dose boosting to improve local control, at least for EC. We also showed that using FLAB improved the overlap between pre- and post-therapy uptakes compared to the use of SUV-thresholds.

(15:15) M5C2-6, Regularization Parameter Selection for Penalized-Likelihood Image Reconstruction in PET

M. Zhang1, J. Zhou2, X. Niu2, C. Ji2, E. Asma2, W. Wang2, J. Qi1

1Department of Biomedical Engineer, UC davis., CA, USA
22Toshiba Medical Research Institute USA, Inc., 706 N Deerpath Dr. Vernon Hills,, IL, USA
Nonquadratic spatial regularizations in iterative PET image reconstruction can preserve edges, but often introduces piece-wise constant blocky artifacts. Hence, an objective, quantitative measurement of blocky artifacts is highly desirable in the design and optimization of penalty functions and regularization parameters in iterative PET reconstruction. Here we propose a scalar metric Q to measure the extent of blocky artifacts in PET images without reference to the true image. The key idea of this Q-metric is based on singular value decomposition of a local image gradient matrix to calculate the locally dominant orientation of textures. The Q-value measured in a relative uniform region in a PET image reflects the strength of artificial edges caused by the blocky artifacts. The proposed Q-metric is applied to real patient images and validated using a human observer study.

(15:30) M5C2-7, Quantitative Image Reconstruction and Detection Performance Evaluation for the Tachyon Time-of-Flight PET Scanner

X. Zhang1, J. Zhou1, Q. Peng2, J. S. Huber2, R. H. Huesman2, W. W. Moses2, J. Qi1

1Department of Biomedical Engineering, University of California, Davis, Davis, CA, USA
2Lawrence Berkeley National Laboratory, Berkeley, CA, USA
The Tachyon is a demonstration single-ring time-of-flight PET scanner that reaches a coincidence timing resolution of 314 ps using LSO scintillator crystals coupled to conventional photo-multiplier-tubes (PMT). The objective of this study is to develop a quantitative image reconstruction method and to quantify the improvement of lesion detection as a result of the high-performance timing resolution of the Tachyon scanner. Detector normalization factors were obtained from a blank scan with a rotating point source. For attenuation correction, we performed post-injection transmission scan with the rotating point source and use the TOF information and rod windowing technique to separate transmission data from emission events. We also implemented a TOF single-scatter-simulation algorithm to estimate the scatters mean. We used an enlarged delayed window to estimate the mean of random events. We examined the image blurring effect by reconstructing point source scans. The results show that our system matrix is accurate enough to model the detector blurring effect caused by crystal penetration. Finally we evaluated lesion detection performance using a combination of separate scans of the NEMA torso phantom and hot spheres of different sizes at different locations. The signal-to-noise ratio (SNR) of a channelized Hotelling observer (CHO) was calculated for each lesion location. To simulate a TOF scanner with 500 ps timing resolution, the measured time difference of each list-mode event was artificially blurred by adding Gaussian random noise. The resulting CHO SNR showed superior performance in lesion detection using the Tachyon scanner compared to a 500 ps TOF PET and a nonTOF PET.

(15:45) M5C2-8, Adaptive Visual-Search Model Observers

H. C. Gifford

Biomedical Engineering, University of Houston, Houston, TX, USA
Realizing the full potential of task-based assessments as a research design tool calls for model observers that can be applied for a variety of clinically realistic tasks. However, existing statistical model observers are task-constrained by the need for extensive prior information. We propose an efficient alternative model that can operate with relatively minimal prior knowledge. For search tasks, our model generates discriminants from feature data containing implicit background information as opposed to available scanning observers that invoke explicit background subtraction. A comparison of the model against human-observer data from a SPECT localization ROC study is presented.