TY - JOUR T1 - Adaptive Autoregressive Model for Reduction of Poisson Noise in Scintigraphic Images JF - Journal of Nuclear Medicine Technology JO - J. Nucl. Med. Technol. SP - 19 LP - 26 DO - 10.2967/jnmt.110.077081 VL - 39 IS - 1 AU - Reijo Takalo AU - Heli Hytti AU - Heimo Ihalainen Y1 - 2011/03/01 UR - http://tech.snmjournals.org/content/39/1/19.abstract N2 - In the present paper, a 2-dimensional adaptive autoregressive filter is proposed for noise reduction in images degraded with Poisson noise. In autoregressive models, each value of an image is regressed on its neighborhood pixel values, called the prediction region. The autoregressive models are linear prediction models that split an image into 2 additive components, a predictable image and a prediction error image. Methods: In this research, unfiltered images were split into smaller blocks, and best combinations of a prediction region and a block size for the image quality of predictable images were sought by using 3 Poisson noise–corrupted images with different image statistics. The images had dimensions of 128 × 128 pixels. Image quality was assessed by means of the mean squared error of the image. The adaptive autoregressive model was fitted into each block separately. Different degrees of overlapping of the image blocks were tested, and for each pixel the mean predictor coefficient of the different models was determined. The prediction error image was calculated for the entire image, and the filtered image was obtained by subtracting the prediction error image from the original image. The effect of the best adaptive autoregressive filter was illustrated using real scintigraphic data. Results: Generally, a prediction region of 4 orthogonal neighbors of the predicted pixel with a block size of 5 × 5 showed the best results. The use of 75% overlapping of the image blocks and 1 iteration of the filtering was found to improve prediction accuracy. The results were further improved when the 2 error term images were summed and subjected to adaptive autoregressive filtering and the resulting predictable image was added to the iteratively filtered image, allowing both noise reduction and edge preservation. Patient data illustrated effective noise reduction. Conclusion: The proposed method provided a convenient way to reduce Poisson noise in scintigraphic images on a pixel-by-pixel basis. ER -