透過直方圖匹配之灰階值超音波影像校正研究-以乳癌超音波影像為例

Abstract

超音波成像提供及時動態影像，有利於放射科專家或醫師即時針對特定影像區域分析並擷取影像。然而，依據操作者經驗、徒手掃描及超音波機器的特性，掃描影像結果的差異可以很大，因此收取的超音波影像缺少比較客觀的標準。傳統上，超音波影像皆是利用直方圖均化 (Histogram Equalization) 或是直方圖匹配 (Histogram Mapping) 方法來校正以獲取較客觀一致的影像，然而當影像內的明暗組織成分組成可以非常不同，若是將明暗成分組成不同的影像校正到固定直方圖下的明暗比例，將無法針對個別影像特性區分欲觀察的組織。 為了針對不同影像特性清楚區分欲觀察的組織，本研究提出標準參考直方圖轉換方法(Reference-Based Histogram Transformation)。首先，尋找每一張影像之標準參考影像區塊，再來，針對此標準參考影像區塊灰階值等比例擴展成灰階值範圍從0到255後，利用此擴展之標準參考影像區塊估計欲轉換的特定分配，再將擴展之標準參考影像區塊進行直方圖匹配(Histogram Mapping)到此特定分配，最後，將直方圖匹配後產生的灰階值轉換對應表應用於整張影像之校正。 本研究以台大醫院提供之266筆乳房腫瘤超音波影像做案例研究，其中包含194筆良性腫瘤和72筆惡性腫瘤，先以本研究之影像校正方法校正每張乳房超音波影像，再取得校正前後影像的乳房腫瘤的量化特徵如輪廓曲折程度量化指標及腫瘤後方回音量化指標，並透過接收者操作特徵曲線(Receiver Operative Characteristic Curve, ROC)，比較影像校正前後的量化指標在判斷乳房腫瘤良惡性的表現。從案例研究結果可驗證，經過校正後的影像量化特徵指標判斷乳房腫瘤良惡性之AUC表現最佳可由校正前的0.679124提升至0.703393471。

Ultrasound imaging provides dynamic images, and it is good for Radiology expert or physician to capture instant images. However, due to machine scan image results affected by operator experience and the limitation of ultrasound image characteristics, the ultrasound image does not have objective criteria. Traditionally, histogram equalization or histogram mapping are used to get the consistency results. Nevertheless, the shading component of organization can be very different. It is difficult to distinguish the individual image organization, if the shading composition of organization is calibrated to the same Histogram. In order to clearly distinguish the organization between different image characteristics, this thesis introduces a Reference-Based Histogram Transformation method. First, we look for the standard reference image block in each image, and then equally scale its original gray level value to interval from 0 to 255. Next, this research adjusts the extended standard reference image block to fit the specific distribution through Histogram Mapping and retrieve a gray level transformation table. Finally, we calibrate the whole image with the gray level transformation table and use the new image as the input of further medical analysis. To evaluate the performance of the novel method, this thesis uses 266 cases (72 malignant lesions and 194 benign lesions) provided by National Taiwan University Hospital (NTUH). This paper introduces the method to calibrate each Breast ultrasound image. Further, quantitative characteristics of breast cancer from the images can be extracted such as quantitative index of Margin Twists and Turns or Posterior Acoustic Shadow. Finally, the research compares the quantitative characteristics of the benign and malignant tumors before and after calibration by Receiver Operative Characteristic Curve (ROC). According to the research, it is proved that the best performance of AUC achieves 0.703393471 after the image calibration by Reference-Based Histogram Transformation. Keywords: Ultrasound Image, Image Calibration, Histogram Equalization, Histogram Matching, Clinical Breast Features, ROC performance.