新冠肺炎斷層掃描的小樣本機器學習研究

Abstract

日前的疫情日趨嚴峻，鑒於目前關於COVID-19之檢測方式大致分為核酸檢測、抗體、抗原之快篩試劑檢測，其分別有耗時高成本高、潛伏期及感染前期無法檢測、以及準確度較低容易發生偽陰偽陽之缺點，且以上採檢方式皆須要透過侵入式的採檢方式才能達到效果，不免會降低社會大眾對於採檢的意願。因此本文希望透過以肺部的電腦斷層攝影圖觀察確診者之病灶位置、形狀、規模並利用奇異值分解對原始資料的主要特徵進行提取，利用圖與圖之間的特徵值差異先進行篩選以及預處理，再利用機器學習中監督式學習的支持向量機及相關的參數設定來對確診者與健康者之間產生區別分類進而對新輸入的病人資料進行預測。透過這種方式期望達到增加檢測量能、降低成本、提高檢測意願及準確度。

The current epidemic is becoming more and more serious, and the current testing methods for COVID-19 are broadly divided into PCR, antibodies, and antigenic rapid test, which are time-consuming and costly, undetectable during the incubation period and pre-infection period, and less accurate which are prone to false negatives and false positives.Also, all of the above methods require invasive testing to achieve results, which will inevitably reduce the public`s desire for testing. Therefore, in this paper, we hope to extract the main features of the original data by observing the location, shape, and size of the lesions of the diagnosed patients with lung computed tomography and using the singular value decomposition, and then filter and pre-process the differences of the feature values between the images. Then we use support vector machine and adjustment the parameters of the model to classify and predict whether the people is diagnosed. In this way, we expect to increase detection performance, reduce costs, and improve willingness and accuracy of detection.