本研究以灰色系統中之灰關聯係數(Grey Relational Coefficient)來發展新的客觀品質評估方法,適當的影像壓縮可在不影響視覺品質下提升儲傳系統之效能。 本研究將Hsia、Wen、Wu、Nagai等人的局部性灰關聯係數(LGRC)進行影像品質評估,藉由CT的Profile與ROI找出最佳的LGRC方法。 實驗以DR、MRI、CT等常見之醫學影像以JJ2000與Apollo進行小波壓縮,壓縮比設定從10倍到100倍,以10倍為一壓縮倍率,並使用MSE、PSNR、SNR、UQI、MSSIM、HVS 與MPR等客觀方法與灰關聯進行比較與討論。 四種LGRC評估方法都能夠正確的評估壓縮影像的品質變化,灰關聯係數隨著壓縮比的增加而降低係數值,其中Nagai的LGRC辨識度較其他三者好,醫學影像的比較結果顯示LGRC與其他客觀方法有相同的趨勢,且JJ2000與Apollo在DR與CT下無任何顯著性差異(p value > 0.05),MRI則是以JJ2000品質較佳(p value < 0.05)。 在評估8張CT影像下,本實驗使用Nagai LGRC僅需0.44秒,比MSSIM的1.2秒與UQI的1.1秒皆快上數倍;在3 × 3之取樣大小時,LGRC之R值依然高達0.999,相較於MRP的0.856好,也略優於MSSIM的0.994與UQI的0.978,所以LGRC不容易受到取樣大小的影響,因此本研究推薦以LGRC來做為醫學影像之評估標準
PURPOSE: The purpose of this study was to develop Grey Relational Coefficient (GRC) system to evaluate compressed medical images. MATERIALS and METHODS: The researcher used Local Grey Relational Coefficients (LGRC) which developed by Hsia, Wen, Wu and Nagai to examine respectively compressed CT, MRI and DR images from compression ratios from 10:1 to 100:1 at 10 different level. The compressed images were also calculated by Mean Squared Error (MSE), Peak Signal to Noise Ratio (PSNR), Signal to Noise Ratio (SNR), Universal Quality Index (UQI), Mean Structural Similarity (MSSIM), Human Visual System(HVS) and Moran''s Peak Ratio (MPR) algorithms. This was to compare the consequences among GRC and the algorithms. The compression algorithms were JJ2000 and Apollo which are wavelet based algorithm. RESULTS: Four LGRCs were all able to evaluate the compressed image quality changes efficiently. The coefficient of GRC decreased with increasing compression ratios. The Nagai method was superior than others in identification. LGRC had similar tendency comparing with other objective methods, JJ2000 and Apollo had no significant difference at 0.05 between DR and CT images, yet compressed MRI with JJ2000 was superior than that compressed with Apollo (p<0.05). For the evaluation speed on compressed images, Nagai''s method was faster than UQI and MSSIM. The second for each image was 0.44,1.2 and 1.1 respectively. For the sampling size of 3×3, LGRC was also superior than MSSIM, UQI and MPR with r-values of 0.999,0.994,0.978 and 0.856 respectively. CONCLUSION: LGRC is not affected by choosing sampling size, LGRC is recommended as for evaluating image quality objectively.