請用此 Handle URI 來引用此文件:
http://tdr.lib.ntu.edu.tw/jspui/handle/123456789/90506
標題: | 基於不同患部影像輸入之深度學習模型應用於非侵入式血紅素濃度偵測 Deep Learning-Based Model for Non-invasive Hemoglobin Estimation via Body Parts Images |
作者: | 林恩廷 En-ting Lin |
指導教授: | 傅立成 Li-Chen Fu |
關鍵字: | 血紅素濃度,貧血偵測,深度學習,電腦視覺, Hemoglobin Estimation,Anemia Detection,Deep Learning,Computer Vision, |
出版年 : | 2023 |
學位: | 碩士 |
摘要: | 貧血一直都是一項全球性的健康議題,根據世界衛生組織(WHO)統計,全球有超過十億人在不同程度上受到貧血問題的影響,就臨床的觀點而言,急性貧血通常是由於出血所造成,嚴重時甚至會危及病人的性命,因此在本研究中,我們著重在對病患血紅素濃度的預測,期望藉由快速且精確的自動化流程來輔助醫師在臨床的診斷。
貧血診斷的黃金標準來自血液中血紅素濃度的實驗室測量,必須通過抽血流程來取得,臨床上為求處置的即時性,醫生經常會透過檢查病人的眼結膜等部位是否蒼白來判斷貧血,然而此方法需要醫生的經驗輔助且具一定的主觀性,因此我們著眼於透過電腦視覺的方法,建立基於眼結膜、舌頭、手掌、甲床四個患部影像輸入的深度學習模型。由於四個患部的影像特徵並不一致,例如在眼結膜影像中微血管的特徵可以提供更多資訊,在其他三個部位卻不容易觀測到此特徵,因此本研究提出了一種新的預測方式,透過輸入額外的患部標籤,搭配融合注意力機制,讓模型在訓練過程中能夠自行學習並強化各個患部的重點特徵,藉以產生足以信賴的結果。與此同時,為了解決資料集中所遇到的資料不平衡問題,我們引入對偶損失函數,讓回歸模型得已受益於廣為使用的分類方法,進而達到穩定處理少數樣本的目的。 總結來說,我們建立了一套基於影像輸入的非侵入式血紅素預測模型,並期望能以現場AI輔助系統的方式為臨床帶來幫助。 Anemia is a significant global health issue, affecting over a billion people worldwide to varying degrees, according to the World Health Organization (WHO). Acute anemia is typically caused by bleeding and can be life-threatening in severe cases. Therefore, this study focuses on the detection of different situations of hemoglobin concentration in patients, intending to assist clinical diagnosis through a rapid and accurate automatic process. Generally, the gold standard for diagnosing anemia relies on laboratory measurements of hemoglobin concentration in the blood, which requires a blood drawing. To meet the need for real-time intervention in clinical practice, physicians often rely on visual examination of specific areas, such as the conjunctiva, to assess pallor and infer anemia. However, this method is subjective and relies on the physician's experience. Therefore, we turn to computer vision techniques and propose a deep learning prediction model based on four input images from different body parts, namely, conjunctiva, tongue, palm, and fingernail. Given that the image features vary across the four body parts, our approach is considered highly novel. By incorporating additional body part labels and employing a fusion attention mechanism, the model learns and enhances the salient features of each body part during training, enabling it to produce reliable results. Furthermore, to address the issue of data imbalance in the dataset, we employ a dual loss function that allows the regression model to benefit from well-established classification methods, thereby achieving stable handling of minority samples. To sum up, we have developed a non-invasive hemoglobin prediction model based on image input, with the goal of supporting clinical practice through an AI-based on-site system. Such results have been verified by real experiment done in National Taiwan University Hospital involving 59 patient subjects, and the prediction accuracy as well as F1-score can achieve as high as 0.658 and 0.778, respectively. |
URI: | http://tdr.lib.ntu.edu.tw/jspui/handle/123456789/90506 |
DOI: | 10.6342/NTU202303357 |
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 | 5.83 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。