RT Journal Article
SR Electronic
T1 <strong>Lesion-Preserving Spatially Adaptive Non-Local Means Post-Filter </strong><strong>for Whole-Body PET Imaging</strong>
JF Journal of Nuclear Medicine
JO J Nucl Med
FD Society of Nuclear Medicine
SP 432
OP 432
VO 61
IS supplement 1
A1 Wenyuan Qi
A1 Chao Song
A1 Chung Chan
A1 Li Yang
A1 Evren Asma
YR 2020
UL http://jnm.snmjournals.org/content/61/supplement_1/432.abstract
AB 432Objectives: PET images suffer from high noise. Non-local means post-filtering has been used successfully for PET noise suppression. In an NLM filter, the image at a given voxel is smoothed by a weighted average of voxels in a large, non-local neighborhood based on the similarities between those voxels and the voxel of interest. Similarity is typically defined as the Euclidean distance between local neighborhoods around voxels and the contribution weight for a surrounding voxel is calculated by scaling this similarity measure. This scaling factor is particularly important as it directly controls the level of smoothing and can have different impacts on regions with different intensity levels. Since whole-body PET images contain voxels with a wide range of intensities, it is challenging for NLM to find a fixed, global smoothing strength that would result in optimal smoothing for every organ. To avoid such scenarios, in this work, we propose a lesion-preserving, spatially adaptive NLM (SA-NLM) filtering approach. Methods: For a given discrete noisy image f, the smoothed result at voxel i from the NLM filter is determined by a weighted average of nearby neighborhood, the weight w(i,j) is based on the similarity between neighborhoods centering at i and j. Image intensity values in the respective local neighborhoods of voxels i and j, denoted by ℵi and ℵj respectively, are used to calculate w(i,j). The difference between the patches is used to define the similarity between the two pixels with an isotropic Gaussian kernel as w(i,j)=exp{(-‖f(ℵi)-f(ℵj)‖2)/h2} where ‖∙‖2 denotes the Gaussian weighted Euclidean distance. The parameter h is the scaling factor that controls the smoothness strength and is a user-selected value. The contribution of a voxel to another voxel’s filtered value is determined by the ratio of ‖f(ℵi)-f(ℵj )‖2 and h2. Therefore, the same h value corresponds to different smoothing strengths if the overall image intensity changes. For regions with lower intensity, the same h will have a larger impact and vice versa. In whole body PET images intensity levels can vary greatly between different organs and patients and this fact means that regions with lower intensity will have higher smoothing while the regions with higher intensity will have less smoothing. This property is not always desirable. In order to overcome this difficulty, we propose a spatially varying strength for NLM where the weights are calculated according to equation w(i,j)=exp{(-‖f(ℵi )-f(ℵj )‖2)/((u(i)*h)2}, where u(i) is the weight map that allows h to vary spatially. In order to obtain the weight map u(i), we apply a median filter on f(i) to remove lesions and salt and pepper-type noise followed by a Gaussian filter to reduce the local variances. This procedure produces a smooth u(i) map that is approximately constant in every organ. We applied this filtering approach on a clinical patient dataset acquired with Canon’s digital TOF-PET/CT scanner with IRB approval and patient’s consent. PET data was reconstructed with fully 3D list-mode TOF ordered-subset expectation-maximization (OSEM) with full physical corrections. In order to quantitatively evaluate the lesion contrast recovery, we inserted two GATE simulated lesions into the acquired clinical data: one in the liver, and one in the lung. Scatter, attenuation and normalization factors were all incorporated into the GATE simulation. Results: In experimental results, we can see that the proposed SA-NLM filter can better preserve lesion contrasts for the lesions that occur in different regions compared to OSEM and constant-strength NLM post-filtering. Conclusions: In this work, we proposed SA-NLM filter with better lesion preservation properties to adaptively smooth different regions of whole-body PET images. Clinical evaluations proved that the proposed method can successfully recover and preserve lesion contrasts in both high background regions such as the liver and low background regions such as the lung.