基於移動窗口和反向殘差互換器之3-D漩渦Transformer於自動乳房超音波腫瘤診斷

Abstract

乳癌是主要造成死亡的癌症之一，及早診斷與治療可以大幅提升乳癌病患的生存率。自動乳房超音波(automated breast ultrasound, ABUS)是常見的乳房檢測技術，提供乳房組織的三維資訊。然而大量二維切片構成的乳房超音波影像讓醫生花費冗長的時間審閱，並且延誤診斷惡性腫瘤。近年來開發了基於深度學習的電腦輔助診斷(computer-aided diagnosis, CADx)系統來自動提取特徵並加速診斷流程。因此本研究提出一個基於Transformer和卷積神經網路(Convolutional Neural Network, CNN)的電腦輔助診斷系統以診斷乳房超音波影像。提出的系統利用自注意力機制來獲得全局影像關係，且加入反向殘差卷積架構來提取局部影像資訊。 本系統包含影像前處理、腫瘤切割和腫瘤分類。在影像前處理，會提取腫瘤區域並調整至固定影像大小，然後利用直方圖均衡化增強影像對比度。接著，調整大小後的影像會透過腫瘤分割模型產生腫瘤遮罩。最後，調整大小後的腫瘤影像、增強後的腫瘤影像和腫瘤遮罩會作為所提出的三維漩渦Transformer腫瘤分類模型的輸入並判斷腫瘤良惡性。而提出的模型是由移動窗口自注意利機制和反向殘差互換器所建構來結合全局和局部特徵。根據實驗結果，本研究提出的電腦輔助診斷系統可達到89.9%的正確率、89.8%的靈敏性、89.9%的特異性和0.9401的曲線下面積。實驗結果顯示提出的系統可以有效輔助醫生更準確的診斷乳癌。

Breast cancer is one of the leading causes of death from cancer. Through early diagnosis and treatment, the survival rate of patients with breast cancer could be significantly improved. Automated breast ultrasound (ABUS) is the most common breast inspection technique and could provide three-dimensional (3-D) spatial information on breast tissue. However, ABUS images consisting of a large number of 2-D slices make radiologists spend much time reviewing and delaying the diagnosis of malignant tumors. Recently, deep learning-based computer-aided diagnosis (CADx) systems have been developed to extract features automatically and speed up the diagnostic process. Thus, this study proposed a CADx system based on the Transformer and convolutional neural network (CNN) for tumor diagnosis on ABUS images. The proposed tumor classification model employed the self-attention mechanism to capture the global relationship and added the inverted residual convolution block to extract local information from images. The CADx system in this study contained data preprocessing, tumor segmentation, and tumor classification. In the data preprocessing, the tumor region was extracted as a volume of interest (VOI) and resized to the fixed image size. Afterward, the histogram equalization was conducted to increase the image contrast. Next, the resized VOI was utilized to generate a tumor mask through the tumor segmentation model. Finally, the resized VOI, the enhanced VOI, and the corresponding tumor mask were used as the inputs of the proposed 3-D SWIRL Transformer tumor classification model to determine tumor malignancy. The proposed model was constructed with shifted window-based self-attention mechanism and inverted residual switch to aggregate global and local features. According to the results, the proposed CADx system could reach 89.9% accuracy, 89.8% sensitivity, 89.9% specificity, and 0.9401 AUC. These results demonstrated that the proposed CADx system could effectively help radiologists diagnose breast cancer more accurately.