利用深度卷積神經網路於乳房磁振造影進行TP53及PIK3CA基因突變預測

Abstract

乳癌是女性最常罹患的癌症之一，若得以在罹癌早期成功偵測，將能降低乳癌死亡率。臨床上經常利用動態磁振造影檢驗患者是否罹患乳癌，而乳癌確診後，若能辨認出有關的生物標記物，將有助於為病患制定一套合適的療程。影像基因圖譜學是新興的研究領域，旨在尋找影像特徵與基因表型間的關聯，提供非侵入式且整體性的量化分析，以此辨識生物標記物。近年來，利用卷積神經網路進行的深度學習方法受到廣泛運用，在包含醫學影像分析等各式各樣的任務中，皆有相當卓越的表現。本論文使用了107筆胸部的三維動態磁振造影樣本，利用深度學習模型自動提取影像特徵，藉此辨認患者的TP53及PIK3CA兩種體細胞基因是否發生突變。在實驗結果中，對於預測TP53及PIK3CA基因突變，準確率分別高達87.97%和82.24%，靈敏性則各自達到82.50%及74.29%，顯示本研究方法能從三維動態磁振造影中有效識別此二種基因突變。

Breast cancer is one of the most common malignancies in females, while an early-stage detection can reduce the mortality. Dynamic contrast-enhanced magnetic resonance imaging (DCE-MRI) is commonly used for breast cancer examination. Once breast cancer is diagnosed, it is more beneficial to recognize related biomarkers for customizing a treatment plan. Radiogenomics research has emerged recently, which focuses on the associations between imaging features and genomic patterns. It provides non-invasive and overall quantification to identify the biomarkers. In recent years, deep learning with convolutional neural networks (CNNs) has been broadly adopted and achieved remarkable performance on various tasks, including medical image analysis. In this study, we propose a method that exploits deep learning to automatically extract image features to differentiate the TP53 and PIK3CA mutations on 3-D breast MRI with 107 cases. The proposed method achieves the accuracies of 87.97% and 82.24%, and the sensitivities of 82.50% and 74.29%, for identifying TP53 and PIK3CA mutations, respectively. The result shows the high potential of this method to recognize these two gene mutations.