請用此 Handle URI 來引用此文件:
http://tdr.lib.ntu.edu.tw/jspui/handle/123456789/90784
標題: | 以注意力強化 ConvNeXt網絡於乳房超音波預測生物標誌 BioCAN: Prediction of Molecular Biomarkers Using Breast Ultrasound Images Based on Attention-augmented ConvNeXt Network |
作者: | 林奕辰 Yi-Chen Lin |
指導教授: | 張瑞峰 Ruey-Feng Chang |
關鍵字: | 乳癌,乳房超音波,電腦輔助診斷,雌激素受體,孕激素受體,人類表皮生長因子受體2,細胞增殖蛋白,卷積神經網絡,注意力機制,自注意力, Breast cancer,Breast ultrasound,Computer-aided diagnosis,Estrogen receptor,Progesterone receptor,Human epidermal growth factor receptor 2,Ki67,Convolutional neural network,Attention mechanism,Self-attention, |
出版年 : | 2023 |
學位: | 碩士 |
摘要: | 乳癌是全球十大癌症死因之一,更是發生率最高的癌症。乳癌可以根據雌激素受體(ER)、孕激素受體(PR)、人類表皮生長因子受體2 (HER2)、與細胞增殖蛋白(Ki67)等生物標記的表達情況,區分為多個亞型,每種亞型的生物學行為、治療反應和治療結果皆不同。目前對生物標記表達的評估依賴於活體組織檢查後的免疫組織化學分析(IHC),這可能無法準確代表整個組織中的目標蛋白和衡量腫瘤時空演變的差異。而研究人員試圖尋找生物標記表達與乳房超音波(US)影像的關聯性,以提供初步且動態的生物行為的監控以協助臨床醫師進行診斷、衡量治療的結果及預後。因此,本研究提出了一種電腦輔助診斷(CAD)系統來預測多種生物標誌物的表達,以協助臨床醫生開立客製化的治療方案並減輕醫務人員的負擔。
我們所提出的系統(BioCAN)包括影像前處理、腫瘤分割和生物標記的預測。首先,我們會提取腫瘤區域並將影像調整為一致的大小。接著,預處理後的影像會透過Swin-Unet分割模型獲得腫瘤遮罩和腫瘤圖像。接下來,我們通過連接原始影像、腫瘤遮罩和腫瘤圖像來獲得融合圖像,以增強腫瘤的內容表達,並將融合圖像作為注意力增強的卷積神經網路(Convolutional neural network, CNN)的輸入來進行生物標誌物表達的預測。我們的分類器以ConvNeXt網路為基底,集成了座標注意力(Coordinate Attention)以及ACmix,為了產生穩定和準確的預測結果,我們還選定準確率前三名的新穎分類器進行加權平均(WA)集成學習。 在本研究中,我們使用總計507張影像評估我們的BioCAN系統,包含了414例ER陽性與93例ER陰性的腫瘤、324例PR陽性與183例PR陰性的腫瘤、83例HER2陽性與424例HER2陰性的腫瘤、以及351例Ki67陽性與156例Ki67陰性的腫瘤。在腫瘤分割中,我們的BioCAN系統可以達到Dice係數為0.9010、IoU值為0.8225、HD95值為17.5166、及ASSD值為6.4369的結果。而在生物標記表達的預測中,我們的BioCAN系統的準確度、靈敏度、特異度和ROC曲線下面積AUC分別在ER狀態的預測達到83.23%、83.57%、81.72%、0.9060;在PR狀態的預測達到77.71%、78.70%、75.96%、0.8526;在HER2狀態的預測達到83.23%、80.72%、83.73%、0.9109,在Ki67狀態的預測達到84.02%、83.76%、84.62%、0.9111。實驗結果證明了本研究提出的方法能夠減輕醫護的負擔並協助放射科醫生預測多種生物標記的表達。 關鍵詞:乳癌、乳房超音波、電腦輔助診斷、雌激素受體、孕激素受體、人類表皮生長因子受體2、細胞增殖蛋白、卷積神經網絡、注意力機制、自注意力 Breast cancer is one of the top ten leading causes of cancer death, with the highest incidence in females worldwide. Breast cancer can be divided into multiple subtypes with different biological behaviors, therapy responses, and treatment outcomes according to the expression of biomarkers, including estrogen receptor (ER), progesterone receptor (PR), human epidermal growth factor receptor 2 (HER2), and proliferation marker Ki67. The assessment of biomarkers expression currently relies on Immunohistochemistry (IHC) analysis after biopsy, which may not precisely represent the target protein in the overall tissue and evaluate differences in the spatiotemporal evolution of the tumor. Recently, researchers have attempted to correlate the expression of biomarkers and breast ultrasound (US), which can provide preliminary and dynamic monitoring of biological behavior to assist the clinician in diagnosing and assessing the outcome of treatment and prognosis. Therefore, this study proposed a computer-aided diagnosis (CAD) system to predict the expression of multiple biomarkers to assist clinicians in prescribing personalized treatment and relieve the burden of medical staff. The proposed system (BioCAN) contains image preprocessing, tumor segmentation, and biomarkers prediction. First, the region of interest (ROI) containing the tumor was extracted and resized to a consistent image size. Afterward, the Swin-Unet segmentation model was adopted to obtain the tumor mask from the ROI. Next, we attained the fused image by concatenating ROI, tumor mask, and tumor image, which emphasize the texture and morphological features of the tumor correlated to the expression of biomarkers. The fused image is then used as the input of the proposed convolutional neural network (CNN) equipped with two attention mechanisms. The proposed classifier integrated the Coordinate Attention module to catch the channel dependencies with direction-sensitive positional information and the ACmix module to introduce the long-range features by mixing the convolution and self-attention paradigm in a single block. Finally, we adopted the weighted average (WA) ensemble learning technique using the models with the Top 3 accuracies to produce a stable and accurate outcome. In this study, we used 507 patients to evaluate our BioCAN system with 414 ER-positive and 93 ER-negative tumors, 324 PR-positive and 183 PR-negative tumors, 83 HER2-positive and 424 HER2-negative tumors, and 351 Ki67-positive and 156 Ki67-negative tumors. The experimental result of the segmentation stage demonstrated that the Dice score coefficient (DSC), Intersection over Union (IoU), Hausdorff distance (HD95), and Average symmetric surface distance (ASSD) were 0.9010, 0.8225, 17.5166, and 6.4369, respectively. In addition, in the biomarkers prediction, the accuracy, sensitivity, specificity, and area under the ROC curve (AUC) of our BioCAN system reached 83.23%, 83.57%, 81.72%, and 0.9060 for ER status, 77.71%, 78.70%, 75.96%, and 0.8526 for PR status, 83.23%, 80.72%, 83.73%, and 0.9109 for HER2 status, 84.02%, 83.76%, 84.62%, and 0.9111 for Ki67 status. The two experimental results proved that the method proposed in this study could reduce the burden of medical staff and assist radiologists in predicting multiple biomarkers expression. Keywords: Breast cancer, Breast ultrasound, Computer-aided diagnosis, Estrogen receptor, Progesterone receptor, Human epidermal growth factor receptor 2, Ki67, Convolutional neural network, Attention mechanism, Self-attention |
URI: | http://tdr.lib.ntu.edu.tw/jspui/handle/123456789/90784 |
DOI: | 10.6342/NTU202300239 |
全文授權: | 同意授權(限校園內公開) |
顯示於系所單位: | 資訊工程學系 |
文件中的檔案:
檔案 | 大小 | 格式 | |
---|---|---|---|
ntu-111-2.pdf 目前未授權公開取用 | 3.3 MB | Adobe PDF | 檢視/開啟 |
系統中的文件,除了特別指名其著作權條款之外,均受到著作權保護,並且保留所有的權利。