應用卷積神經網路於分析動態增強核磁共振影像之乳癌生物標記

DCE-MRI Breast Tumor Biomarker Analysis Using Convolutional Neural Network

徐宏毅(Hung-Yi Hsu)

指導教授 ： 張瑞峰

國立臺灣大學/電機資訊學院/資訊工程學研究所/碩士(2017年)

https://doi.org/10.6342/NTU201703493

摘要

根據研究顯示乳癌目前仍然高居女性前十大死亡原因之一。而近年來，醫生在治療乳癌病人時會根據生物標記來擬訂不同的治療策略。而常見的生物標記有動情激素受體受體(ER)、黃體激素受體(PR)、第二型人類表皮生長因子受體(HER2)。研究顯示，這三種受體對於乳癌病人的治療方式選擇有顯著的影響。當三種受體同時為陰性時，此時會特別稱呼此為三陰性乳癌。三陰性乳癌對於藥物治療的療效都較差，因此會特別將之當成一種類別判斷。在這篇論文中，我們應用在卷積神經網路在動態增強核磁共振影像判斷生物標記種類，並結合之前研究所提取的人工特徵。動態增強核磁共振影像是在病人注入顯影劑後在不同時間點所拍攝的一系列影像，藉由顯影劑能清楚顯現腫瘤區域。在這篇論文中我們提出了一個不同於一般的卷積神經網路模型，此種模型除了應用影像資料，並增加人工提取特徵用來預測生物標記，在這篇論文中我們稱之為混合模型，共有13層不同類型的應用神經網路層。混和模型的輸入資料結合了動態增強核磁共振影像與傳統方法為人工思考影像之間的相關性再利用電腦所提取的影像特徵值，而模型的組成類似於一般卷積神經網路模型但在模型中額外加入了人工提取特徵，此種方法增加了多樣性的特徵並用於訓練卷積神經網路模型來增加準確度。整體來說，混合模型在預測四種類型生物標記的準確率上有最好的表現，特別是在第二型人類表皮生長因子受體。最後的結果顯示，混合模型在動情激素受體分類實驗中，102個實驗資料裡包含59位陽性反應與43位陰性反應，結果達到準確性74.5% (76/102)、敏感性79.7% (47/59)、專一性67.4% (29/43)以及Az值0.7382。在黃體激素受體分類實驗中，102個實驗資料裡包含38位陽性反應與64位陰性反應，結果達到準確性72.5% (74/102)、敏感性47.3% (18/38)、專一性87.5% (56/64)以及Az值0.6472。而在第二型人類表皮生長因子受體分類實驗中，102個實驗資料裡包含47位陽性反應與55位陰性反應，結果達到準確性84.3% (86/102)、敏感性85.1% (40/47)、專一性83.6% (46/55)以及Az值0.8492。最後，在腫瘤的三陰性分類實驗中，總共有102個病理檢驗過的腫瘤作為實驗資料，其中包含80個有三陰性反應的腫瘤以及22個無三陰性反應的腫瘤，最後的結果能達到78.4% (80/102)、敏感性45.5% (10/22)、專一性87.5% (70/80)以及Az值0.7098。最後，根據我們的結果顯示，在結合了卷積神經網路模型與人工提取特徵後，能在之前的研究結果上更進一步提高準確度。

關鍵詞：核磁共振影像、乳房、動情激素受體、黃體激素受體、第二型人類表皮生長因子受體、三陰性、卷積神經網路、人工提取特徵

關鍵字

核磁共振影像 ； 乳房 ； 動情激素受體 ； 黃體激素受體 ； 第二型人類表皮生長因子受體 ； 三陰性 ； 卷積神經網路 ； 人工提取特徵

並列摘要

Breast cancer is the women leading cause of the death. In the recent years, doctors will develop different treatment plan according to molecular biomarkers of breast cancer. There are several biomarkers which were used in treatment strategy and prognosis estimation, such as estrogen receptor (ER), progesterone receptor (PR), and human epidermal growth factor receptor 2 (HER2). Breast cancer expresses negative in ER, PR, and HER2 is known as triple negative breast cancer (TNBC). TNBC had bad efficacy in medicine therapy. In this paper, the combination of handcraft feature and feature maps extracted by Convolutional Neural Network (CNN) is used to predict the statuses of ER, PR, HER2, and TNBC on breast dynamic contrast-enhance magnetic resonance imaging (DCE-MRI). We developed a different model derived from general CNN model to predict biomarkers of tumors. This model not only used image data as input but jointed handcraft features into CNN model, named as the combined model. The combined model which used both image data and handcraft features as input providing diversity of features in CNN model training. Therefore, we got better accuracy on molecular biomarker prediction. Generally, the combined model had the best performance in accuracy for HER2 classification. In the experiment of ER classification, 102 biopsy-proved tumors with 59 ER positive tumors and 43 ER negative tumors were used to evaluate the performance of the combined models. Its accuracy, sensitivity, specificity, and Az value are up to 74.5% (76/102), 79.7% (47/59), 67.4% (29/43), and 0.7382. For PR classification, there are 102 biopsy-proved tumors with 38 PR positive tumors and 64 PR negative tumors. Its accuracy, sensitivity, specificity, and Az value are up to 72.5% (74/102), 47.3% (18/38), 87.5% (56/64), and 0.6472. For HER2 classification, there are 102 biopsy-proved tumors with 38 HER2 positive tumors and 64 HER2 negative tumors. Its accuracy, sensitivity, specificity, and Az value are up to 84.3% (86/102), 85.1% (40/47), 83.6% (46/55), and 0.8492. For TNBC classification, there are 102 biopsy-proved tumors with 22 triple negative tumors and 80 non-triple negative tumors. Its accuracy, sensitivity, specificity, and Az value are up to 78.4% (80/102), 45.5% (10/22), 87.5% (70/80), and 0.7098. In summary, our combined model, which combined CNN and handcrafts features, improved the performance of identifying molecular biomarkers in breast cancer.

Keywords: DCE-MRI, breast, estrogen receptor, progesterone receptor, human epidermal growth factor receptor 2, Triple negative, CNN, handcraft feature

並列關鍵字

DCE-MRI ； breast ； estrogen receptor ； progesterone receptor ； human epidermal growth factor receptor 2 ； Triple negative ； CNN ； handcraft feature

參考文獻

[1] R. L. Siegel, K. D. Miller, and A. Jemal, "Cancer statistics, 2016," CA: A Cancer Journal for Clinicians, vol. 66, no. 1, pp. 7-30, 2016.

[2] G. E. B. C. T. Collaborative, "Effects of chemotherapy and hormonal therapy for early breast cancer on recurrence and 15-year survival: an overview of the randomised trials," The Lancet, vol. 365, no. 9472, pp. 1687-1717, 2005/05/14/ 2005.

[3] M. C. U. Cheang et al., "Basal-Like Breast Cancer Defined by Five Biomarkers Has Superior Prognostic Value than Triple-Negative Phenotype," Clinical Cancer Research, vol. 14, pp. 1368-1376, 2008-03-03 11:08:36 2008.

[4] T. C. Putti et al., "Estrogen receptor-negative breast carcinomas: a review of morphology and immunophenotypical analysis," Mod Pathol, vol. 18, no. 1, pp. 26-35, 08/27/online 2004.

[6] E. H. Romond et al., "Trastuzumab plus Adjuvant Chemotherapy for Operable HER2-Positive Breast Cancer," New England Journal of Medicine, vol. 353, no. 16, pp. 1673-1684, 2005.