Generalization evaluation of numerical observers.p

文件名称: Generalization evaluation of numerical observers.pdf

所属分类: 医疗

开发工具:

文件大小: 192kb

下载次数: 0

上传时间: 2019-10-14

提供者: crono*****

下载 (192kb)

不能下载？报告错误

详细说明：numerical observers比较好的一篇文献，比较好的一篇文献Figure 2(left)shows the calculated A, by the different generalize better when the test images are reconstructed in a methods. These results suggest that both CSVM and CHO can different way from that of training images. Therefore, the accurately generalize to unseen images, provided that these CSVM method could be more adequate for evaluating images are produced in exactly the same way as those used in c reconstructed images for which no human observer data are training. available Comparison 2. Generalization from one broad class of images to another IV. REFERENCES In this comparison, we studied a form of generalization [1] H.H. Barrett and K Myers, Foundations of Image Science, New York that is perhaps most representative of the practical use of a Wiley, 2003, Chap. 14 numerical observer. We trained both CSVM and CHO on a [2] K.J. Myers and HH. Barrett, J. Opt. Soc. Am. A, vol 4, no 12,pp 2447-2457,1987 broad range of images, and tested them on a different, but [3] J. Yao and H H. Barrctt. Proc. SPIE, 1786, pp 161-168, 1992 equally broad, set of images. Specifically, we trained both [4] S.D. Wollenweber, et al. Proc. IEEE Nucl. Sci. Symp., vol 3, pp 2090 numerical observers using all the filter FWhm values and 2094,1998. one-iteration OSEM, and then tested the observers using all [5 H.C. Gifford at al. IEEE Trans. Nucl. Sci., vol. 46. no. 4,, pp. 1032 the filter fWhm valucs and fivc-itcration OSem 1037,1999 The results of this experiment are shown in Figure 2 [6] K abbey and H.H.Barrett,J.Opt.Soc.Am. A, vol 18,no. 3, pp 473 4882001 (right). In this situation, the Cho performed relatively poorly, [7] M. V. Narayanan et al. IEEE Trans. Nucl. Sci. vol. 49, no 5, pp 2355 failing to match either the shape or amplitude of the human 2360,2002 observer curves, while the CSvM was able to produce [8] J. Oldan et al., IEEE Trans. Nuc. Sci, vol. 51, pp. 733-741, 2004 [9 P. Bonetto et al., IEEE Trans. Nucl. Sci, voL 47, no 4, pp. 1567-1572 reasonably accurate predictions 2000 In all experiment, the parameters of the CHo and CSVM [10] T K Narayan and G T Herman, J. Opt. SoC. Am. A, vol. 16, no. 3 were optimized to minimize generalization error measured 1999 using five-fold cross validation based on the training images [I N Cristianini and J. Shawe-Taylor, Cambridge: Cambridge Univ Press, only [12]MN. Wernick, J. Opt. Soc. Am. A, vol 8, pp 1874-1880, 1991 3 J.G. Brankov ct al., IEEE Nucl Sci. Symp. Mcd Imag. Conf, vol 4 IIL CONCLUSION pp.2526-2529,2003 L4」 ROCKIT, In this paper we developed and evaluated a NO using a http://xray.bsd.uchicago.edu/krl/krlRocSoftwaReindexhtm channelized SVM. Our results demonstrate that while both [15] I.M. Iludson and R.S. Larkin, IEEE Trans. Med Imaging, vol 4, pp CHO and CSVM can generalize well when training and test [16] Narayanan, M.V. et al. voL 49, no 5, pp.2355-2360,2002 images are both reconstructed in the same way, the csvm can 5 ite rati 5 ite ration 95 0.95 0.9 0.85 085 0.8 0.75 CSVM CHO 07 0.7 HO 0.65 0.65 5 FWHM (pixels) FWHM (pi eels) 1, l Figure 2. Calculated area under the ROC curve for HO, CSVM and CHO, where error bars represent plus/minus one standard deviation. Left: Comparison 1698

(系统自动生成,下载前可以参看下载内容)