融入乳房解剖資訊之全乳房超音波影像腫瘤偵測演算法

Abstract

全乳房超音波為新一代的乳癌篩檢工具，為了使此工具能更廣泛的應用，並協助醫師偵測與診斷乳房腫瘤，避免遺漏可疑的病灶，我們提出一全乳房超音波影像腫瘤偵測演算法。首先對影像做一個前景遮罩，將未照到乳房的部分去除，再對遮罩內的乳房影像做四組不同範圍的multi-scale球狀偵測，利用接收超音波訊號信心程度與乳房解剖資訊，大幅將非腫瘤的球狀結構刪除後，聯集四組不同範圍所剩餘的球狀結構為一組，再針對聯集後剩餘的球狀結構，擷取球狀結構的類球狀機率、大小以及位於肌肉層的機率特徵，使用MIFS(mutual information based feature selection)，挑選前三個有效的特徵，以邏輯斯迴歸分析模型作為分類器，進行LOOCV交互驗證的方法，並實作Moon等人的研究，使用相同的影像，以及相同的交互驗證方法與我們的演算法比較，使用29位病人，共49組影像，包含86個病灶，得到ROC曲線下面積0.93高於Moon等人方法0.82的結果。

Whole breast ultrasound is a new generation of screening tool for breast cancer. In order to make this new technology more widely applicable, assist lesion detection and diagnosis in clinical practice, and avoid the missing of potential lesions, we propose a lesion detection algorithm for whole breast ultrasound. We firstly conduct a masking preprocess to discard the dark non-breast regions in the image volume, and then apply four passes of multi-scale blob detection on the remaining regions. Following that we discard most unlikely blob structures according to a computed confidence map and the prior knowledge of breast anatomy. The confidence map is derived from the physical ultrasonic property. After this discard step, we collect the survival blob structures from the results of four passes blob detection in a single set. The features of blobness, size, and probability of being at muscle layer of the all survival blob structures are adopted in a classification process for the differentiation of true lesions from negative ones in the collection set. MIFS (mutual information based feature selection) procedure is applied to select three most effective features for the purpose of dimension reduction. With the aid of logistic regression classifier and the process of LOOCV cross validation method, we are able to achieve high ROC area in 0.93 on the testing 49 image volumes. The 49 image volumes were acquired from 29 patients and contains totally 86 lesions. Moon's method is also implemented as baseline with the performance of ROC area in 0.82. The experimental results suggest that our method is better than Moon's lesion detection method.