上消化道診斷之超解析度放大技術之開發

Development of image super-resolution technology for upper gastrointestinal diagnosis

張孝慈(Hsiao-Tzu Chang)

指導教授 ： 蘇振隆

中原大學/工學院/生物醫學工程研究所/碩士(2015年)

https://doi.org/10.6840/CYCU.2015.00142

摘要

窄帶光譜影像和Fuji 的內視鏡影像系統都是目前腸胃診斷時常使用的影像增強技術方法。影像超解析度的方法，藉由影像放大可以提升影像品質，有利於醫生在上消化道的早期癌症的診斷。本研究是基於使用計算機視覺與色彩空間轉換的檢測技術而檢測出癌症的病變區域之概念，而進行之內視鏡用在上消化道影像的超解析度演算法之開發
在研究中首先將原始影像放大成原來的4倍大小，並將影像分成高頻與低頻影像進行處理，藉由調整Alpha值控制影像中高頻部分，讓影像中的輪廓以及邊緣部分更清晰。最後結果使用20組影像來測試其效果，每一組影像有：鄰近內插法、線性內插法、雙立方內插法、Neighbor Embedding、Sparse Recovery以及本研究方法進行比較。在分析與評估部份分為主觀與客觀兩個部分進行。客觀部分是除將放大後影像與原始影像做相減，並將相減後的數值進行方均根與其他方法相比外；亦計算其運算速度。而主觀的部份是將影像做評估問卷調查，評分最高為7分，並請1位專科醫師5位學生幫忙評定。
結果顯示：本方法之Alpha 值設定在1.05至1.15為之間可得最佳的結果，以256*256為例本研究其方均根差為1.186，相較於次佳的雙三次內插法為1.264以及鄰近內插法為1.289，本研究結果的誤差數值為最小，差異最小；在時間部分，與Neighbor Embedding、Sparse Recovery 等方法相比較後，本研究處理時間為4.268 秒，相較於Neighbor Embedding 462.386秒以及Sparse Recovery 275.176秒來的少，在時間處理的比較上相對的處理時間是較快的。而本研究經由統計結果得到的平均分數為5.12分，相較於Sparse Recovery 方法的4.99分以及雙三次內插方法4.59分為高。
總體而言，本研究所提供之方法可藉由使用者的喜好度，調整影像高頻部分，以達到使用者最想要的最佳結果。並由數據可知本研究最後顯示的影像結果，相對於其他方法顯示是可得到較好的結果。
能讓醫師在觀察內視鏡時，更有效的確認病灶，以提供醫師更有效的診斷。

關鍵字

超解析度 ； 上消化道內視鏡影像診斷之分析 ； 內視鏡影像分析 ； FICE內視鏡影像增強技術

並列摘要

Both narrow band spectrum imaging and Fuji endoscopy imaging systems are usually used to enhance image for diagnostic intestines and stomach. The super resolution imaging method which elevates quality of image by magnifying image are used to benefits doctors to diagnostic the cancer in early stage in the digestive tract. This research is based on the concept of transformation of vision and colors space technique to detect the area of cancer and further proceeds in development of super resolution algorithm by using endoscopy on the upper digestive tract.
In this study, image is divided into high and low frequency and was processed after magnifying four times original imaging. By regulating Alpha value to control high frequency component of image can made the outline and fringe of image clearer. The research result was compared with methods of the nearest neighbor interpolation, linear interpolation, bi-cubic interpolation, Neighbor Embedding, and Sparse Recovery by applied 20 exist images. Both subjective and objective tests are used to analysis and evaluate the processed image. For objective evaluation, the root mean square error method was used and the time cost for different methods also calculated. A questionnaire survey of imaging evaluation judged by a diplomate and 5 students and 7 scores system was used for subjective evaluation.
The result indicated that the best result was acquired by setting Alpha values in the interval between 1.05 and 1.15 of this method. Based on 256*256 images, the root mean square error of this research is 1.186. Compared with 1.264 for bi-cubic interpolation method and 1.289 for nearest neighbor interpolation, the error of this method is the least and is of the least variation as well. For temporal part, processing time of this method is 4.268 seconds which was less than 462.386 seconds for Neighbor Embedding method and 275.176 seconds for Sparse Recovery method. For 7 scoring system in subjective evaluation, the average score is 5.12 for this developed method which higher than 4.99 scores for Sparse Recovery and 4.59 scores for bi-cubic interpolation in this study.
In summary, this research can regulate high frequency of image by users’ preference to reach the best result that the users want. Better results compared with other methods were indicated through data which indicated last result of image. Furthermore, doctors can confirm nidus and diagnostic more efficiently when they are using endoscopy.

並列關鍵字

Endoscopic image analysis ； FICE endoscopic image enhancement technology ； Super-Resolution

參考文獻

[12] 廖振偉、史天元、張崑宗, "細化演算法比較," 中華民國地籍測量學會會刊,vol. 20, pp. 1-18, 2001.

[17] 孫善群, 利用頻譜分析胃腸道系統影像之輔助診斷系統, 原大學生物醫學工程學系碩士論文,中壢, 2013.

[18] 廖榮聖, 無線膠囊內視鏡影像資訊系統之設計, 中原大學生物醫學工程學系碩士論文, 2003.

[2] M. H. Shengyang Dai, Ying Wu, Yihong Gong, "Bilateral Back-Projection for Single Image Super Resolution," Multimedia and Expo,pp. 1039-1042 , 2007.

[4] S. S. K.Ngyuyen, C.Fookes, "Feature-domain super-resolution framework for Gabor-based face and iris recognition," Computer Vision and Pattern Recognition (CVPR), pp. 2642-2649, 2012.