We devised an adaptive partial smoothing filter (APSF), which is a specially designed filter used to perform local smoothing using a variable filter size and shape determined by the distribution of pixel values of contours or edges in the region of interest. By adjusting 3 major parameters that exist in the APSF, an optimal condition for image enhancement can be obtained.
1. Description of the APSF
The basic algorithm for filter size and shape determination in the APSF was based on a method reported by Guis et al [4]. The outline of the APSF is as follows.
1) After applying a 55 averaging filter to the original image, each pixel (i, j) of the image I is assigned an upper window Wmax centered on it, whose size is NN .
2) Let I(i, j) be the pixel value of pixel (i, j) in image I, and let T be a given threshold. Pixel (k,l) within Wmax is assigned a binary mask value 0 if I(k, l) I(i, j) >T, else it is assigned a binary mask value 1. This results in constructing a binary image (Fig. 2) .
3) For each window size C C [C=3,5,...,.N], the percentage P0 of zeros is computed over the region of external areas ( red ). Let C0 be the maximal value beyond which the percentage P0 isgreater than 60%. The pixel (i, j) is assigned the window W = C0 C0 (Fig. 3) .
4) Finally, the processed image I' is computed from
I'(i, j) = M (i, j),
where M (i, j) is the mean value in image I of pixels labeled as a binary mask value 1 in the window C C around pixel (i, j) (Fig. 4) .
It is noted that the processed image I' largely varies depending on the threshold value T .
2. Composite images
The aim of this simulation was to determine the optimal threshold value, which was used in applying the APSF to clinical images. A composite image was obtained by adding a computer-simulated lentiform-nucleus like object (LN) to a uniform phantom image generated from 4 slice CT scanner (Fig. 5) ,
3. Performance Evaluation
Two criteria were used to evaluate the performance of the APSF on the composite images. From the evaluation results, an optimal threshold value used for clinical images was obtained.
1) Noise reduction
The standard deviation (SD) of pixel values in an image is used to quantify noise reduction. Low SD value implies that high reduction of noise can be obtained by the APSF. The SD of the pixel values was obtained from a region of interest (ROI) in the composite images (Fig.6) . The SDs were computed by changing the threshold value T within the range [1.5, 5.0].
2) Edge blurring
To investigate the extent of edge blurring, slope ratio (SR) was calculated from a profile of pixel values measured at the right angles to the edge of LN. The SR is defined as
SR(%) = Ppr / Porg 100,
where Ppr and Porg are the pixel-value difference of between 2 pixels at 3-pixel distance apart for the processed image and that for the original image, respectively (Fig. 7) . The 3-pixel distance corresponds to the length of the edge profile for the images. Low SR value means that high edge blurring occurs resulting from the APSF. The SRs were evaluated by changing the threshold value T within the range [1.5, 5.0].