基於深度學習技術之間質性肺病紋理分類

間質性肺病（Interstitial Lung Disease, ILD）為肺炎的一種，多數的間質肺病末期時會產生肺纖維化，爾後引起肺部僵硬，降低氣囊攜帶氧氣進入血液之能力，最終導致永久喪失呼吸能力、心肺功能衰竭並死亡。且台灣自民國100年來，肺炎死亡人數逐年增加，且65歲以上長者就有多達11365名，人數比死於肺癌的還多。由此可知年長者是感染肺炎的高危險群，提早預測並加以治療是必要的。本論文以深度學習之卷積神經網路進行間質性肺病紋理之分類，使用NasNet、VGG16、DenseNet及Xception卷積神經網路架構來訓練影像，搭配四種激勵函數，四種優化器訓練並比較其一百一十二種搭配組合之優劣，藉此找到VGG16之FC1搭配ReLU激勵函數、FC2搭配ReLU激勵函數及搭配Adam優化器，命名為RRANet-VGG16，達到Metrics Accuracy為99.73%、Metrics Precision為99.73%、Metrics Recall為99.73%及Metrics F1 Score為99.73%，RRAnet-VGG16為一個可用於辨識間質性肺病紋理的最佳卷積神經網路架構。以輔助胸腔科醫師減少判讀間質性肺病紋理影像的時間並提高其診斷之準確性，使病患提早治療，藉此以提高病患之存活率。

關鍵字

間質性肺病；深度學習；卷積神經網路；影像辨識

並列摘要

Interstitial Lung Diseases (ILD) is a type of pneumonia. Most of the interstitial lung diseases produce pulmonary fibrosis at the end of the period, which causes lung stiffness and reduces the ability of the balloon to carry oxygen into the bloodstream, eventually leading to permanent loss of respiratory capacity, cardiopulmonary failure and death. Moreover, since the Republic of China in the past 100 years, the number of pneumonia deaths has increased year by year, and there are as many as 11,365 elderly people over the age of 65, more than the number of people who died of lung cancer. It can be seen that the elderly are a high risk group for pneumonia, and it is necessary to predict and treat them early. This thesis uses the deep learning convolutional neural network to classify the texture of interstitial lung disease, using NasNet, VGG16, DenseNet and Xception convolutional neural network architecture to train images, with four activation functions, four optimizers training and compare the advantages and disadvantages of the one hundred and twelve combinations, to find the FC1 with the ReLU activation function of VGG16, the FC2 with the ReLU activation function and the Adam optimizer, named RRANet-VGG16, and the Metrics Accuracy is 99.73%, Metrics Precision is 99.73%, Metrics Recall is 99.73%, and Metrics F1 Score is 99.73%. RRANet-VGG16 is an optimal convolutional neural network architecture that can be used to identify interstitial lung disease texture. To help the thoracic surgeon reduce the time to interpret the texture image of interstitial lung disease and improve the accuracy of its diagnosis, so that patients can be treated early to improve the survival rate of patients.