腫瘤超音波影像判別系統分析、設計與建置

Abstract

超音波檢查為早期腫瘤最有效的非侵入性篩檢方法之一，專業醫師根據所觀察之影像特徵提出持續追蹤或進一步檢查之建議。由於目前超音波影像皆為經由醫師主觀擷取並判斷，不同觀察者對於相同影像常有不同解讀而導致差異明顯的檢查結果，因此影像分類之客觀化成為重要課題。因此本研究希望能透過影像量化的方式提供一個診斷輔助系統,提供客觀且可靠的資訊輔助醫師做臨床診斷。但此舉動將會面臨兩大難題:第一,如果將超音波影像量化,並利用數學模型對影像量化後所得之特徵值執行一系列演算,計算出一個較為客觀的腫瘤良惡性結果供醫師做為診斷參考,此過程常相當複雜且耗時。第二,建立一個用來區分腫瘤良惡性之數學模型往往同樣的非常耗時且複雜,又當未來有新訓練資料加入時,需要花相當多的人力與時程去重建分類模型。因此本研究針對上述第一難題,整合了複雜的模型分類過程,建構一個簡易快速的平台系統提供客觀的標準答案輔助醫生做腫瘤診斷,另外針對第二難題提供工程師一個可簡易操作的模型建置精靈(Wizard),來節省建立模型的所耗費的心力。 本研究中系統背後用來執行腫瘤分類的模型是利用莊曙詮學者在2012所提出的多階段調適樹群模型,而此樹群模型的建構中所使用的超音波腫瘤影像指標則是楊紹桓學者在2012所提出。本系統可概分為兩個模組,第一個模組是用來輔助醫生臨床診斷，能提供腫瘤良惡性的參考答案給醫生作為參考,本研究按照系統分析方法,遵循系統生命週期去分析、設計,將莊曙詮學者複雜的模型分類過程整合到醫師用來檢視超音波影像的系統中,提供腫瘤良惡性的參考答案來輔助醫師做臨床診斷。而第二個模組是設計來幫助工程師,建立輔助診斷系統背後用來判斷腫瘤良惡性的分類模型，提供一個友善的建置精靈(Wizard)給工程師一步步的建立出此分類模型，縮短工程師在建立模型上所要花的時間。因為在莊曙詮學者論文中所提出的分類模型建置工作包含複雜的三階段程序,需要用人力去執行一系列高重覆性的的建置工作,如果由工程師手動執行需要耗費大量時間且容易出現人為錯誤。故在本研究中按照了系統分析的方法,遵循系統生命週期去分析、設計,並自動化了建置模型背後的複雜工作,使工程師不再需要以人力去執行一系列複雜的模型建置過程。

本研究最後利用臺大醫院所提供的264筆乳房腫瘤與433筆甲狀腺腫瘤的樣本資料來進行此資訊系統測試以驗收其在分級系統(BI-RADS/TI-RADS)的分類效能,確保可信度與可靠度。

Ultrasound (US) imaging is one of the most effective non-invasive screening tools for tumors of early stage. Based on observation impressions of US images, clinicians make suggestions for patients to be subject to periodic follow-up or further cytologic tests. Because acquisition and observation of ultrasound images are mostly subjective and highly dependent on the medical staff’s experience and judgment, the observer variation often results in significantly different decisions. Objective quantification of sonographic tumor features has become a pressing issue facing the medical staff. This research focuses on constructing a tumor classification system that provides straightforward, convenient working bench for computer-assisted-diagnosis (CADXx) engineers to perform computerized quantification of sonographic features and to establish tumor classification models. The system will also provide a friendly interface to assist doctors making their diagnosis baseon the classification models. The classification model used in this research is “Adaptable Multi-phase Ensemble model” (Shu-Chuan Chuang, 2012) while the methods to quantify the sonographic feature of tumors in constructing the model are given by Shao-Huan Yang (2012). The proposed tumor classification system consists of two main modules. The first module is to support CADXx engineers in constructing the mathematical models used to classify tumors based on the quantified sonographic features. This module provides a step-by-step interface for engineers to construct the mathematical model. In addition, an engineer can perform sensitivity analysis with varied model parameters. This proposed module will significantly shorten the time required to construct a complicated classification model and allow the built models stored in the system for future use. The second module is then designed to help doctors in practical sonographic tumor diagnosis. This module aims to assist doctors in making their tumor diagnosis by providing automatic tumor margin segmentation, sonographic features quantification, the probability of tumor classification and the summary of critical sonographic features. To validate the system performance, we use a database of 264 cases of breast tumors and 433 cases of thyroid tumors provided by National Taiwan University Hospital (NTUH) to build BI-RADS and TI-RADS classification models, respectively, and to test the system’s standalone performance of tumor classification.