Segmentation of Nuclear Medicine three-dimensional images using Anscombe transformation

In general, Nuclear Medicine images present poor signal to noise ratio mostly due to low counts and Poisson noise. The success of segmentation may be improved if relevant information of the input image is enhanced. Our aim was to verify segmentation improvement due to adequate filters. Basically, all images were submitted to Anscombe Transform, filtered by using Lee filter and inverse Anscombe Transform. Then, they were segmented by using Fuzzy Connected method and finally, compared to expected results. This study simulated real images obtained in Nuclear Medicine, such as the volume of the left ventricle of the heart (128×128×128 voxels), then Poisson quantic noise was incorporated in the simulation. For the assessment we used 30 three-dimensional images with different dimensions (volume, wall thickness) and different instances of Poisson noise, divided in three groups: Group I - 10 three-dimensional images of adult males, Group II - 10 three-dimensional images of adult females and Group III - 10 three-dimensional images of children. The analysis used the following parameters: True Positive (TP) rate and False Positive (FP) rate, and Maximum Distance (MaxDist) to the expected contour. Our results show consolidated results obtained from the validation of the image processing, by highlighting the usefulness of Anscombe/Lee filter, and showing the superiority of the segmentation values due to the proposed process.