利用卷積神經網路建立褥瘡傷口分析系統

Abstract

褥瘡是長期護理或醫院護理中的常見問題。褥瘡是軟組織長時間受壓引起的，會造成局部組織損傷，甚至導致嚴重感染。褥瘡可能導致預後不良、長期住院和增加醫療費用，這在老齡化社會中尤其成問題。褥瘡的治療目標需要根據其不同階段和感染程度進行治療。治療選擇包括減少風險因素、治療局部傷口以及在必要時服用抗生素。本研究主要採用深度學習的方法對褥瘡進行分析、診斷和輔助決策，為一線護理及照顧人員提供更多參考。 我們所建立的褥瘡傷口診斷以及輔助決策系統主要包含傷口分析系統以及資料庫管理系統。其中傷口分析系統包含：一、自動傷口輪廓分割，二、自動傷口大小測量，三、自動傷口診斷和治療建議。 為了建立自動傷口輪廓分割，我們訓練比較兩種不同之深度學習架構 : （1）語意分割U型全卷積神經網路（U-Net），（2）實例分割遮罩型區域卷積神經網路。 為了實現自動傷口大小測量，我們採用了雷射雷達技術（Light Detection and Ranging，雷射探測與測距）找出空間深度以及三維座標，同時合併前項訓練良好的分割法找出傷口輪廓座標進而算出傷口面積。 至於自動傷口診斷和治療建議，我們提出了根據目前最新褥瘡治療準則所改良的判斷流程，其中包含了兩個深度學習的分類任務：發紅度分類與壞死度分類。我們訓練並比較四個經典的卷積神經網路架構：AlexNet，VGG，ResNet以及Inception-ResNet-V2。 我們分析比較以上三個任務和多種不同的深度學習架構，找出最佳表現的神經網路模型，以建立完善我們的褥瘡傷口診斷以及輔助決策系統。在自動傷口分割上，U型全卷積神經網路的表現勝於遮罩型區域卷積神經網路（交聯比: 0.7773 對0.4604）。在自動傷口大小測量上，我們對傷口面積的評估平均相對誤差為26.2%。而至於發紅度分類與壞死度分類，Inception-ResNet-V2有最好的表現，兩者分別達到98.5%以及97%的準確率。 在研究測試上，褥瘡傷口診斷以及輔助決策系統有著不錯的表現，但在臨床使用端還有使用者回饋的部分，還有待前瞻型研究實測驗證，以期能真正改善臨床執業的環境，達到真正輔助一線臨床工作者的目的。

At nursing homes or hospitals, pressure injuries are a frequent problem. Long-term compression of soft tissues results in pressure injuries, which can harm nearby tissue and possibly spread dangerous infections. Pressure injuries may result in poor prognosis, prolonged hospitalization, and higher medical expenses, all of which are problems in an aging population. The treatment goals of pressure injuries need to be treated according to their different stages and the degree of infection. The treatment option includes reducing the risk factors, treating the local wounds, and taking antibiotics if necessary. This study mainly uses the methods of deep learning to assess and assist making decision for pressure injuries and provide more references for first-line caregivers. Our Pressure-Injury Assessment System (PIAS) is composed of wound analyzation system and database management system. The wound analyzation system was composed by three main components: 1. Automatic wound segmentation, 2. Automatic wound area measurement and 3. Automatic wound diagnosis and treatment suggestion. For automatic wound segmentation, we trained and compared two deep learning models: 1. Semantic segmentation: U-Net and 2. Instance segmentation: Mask Region-based Convolution Neural Network, Mask-RCNN. For automatic wound area measurement, we adopted the LiDAR technology (Light Detection And Ranging) to find the spatial depth and 3-dimentional (3D) coordinates. Combined with the well-trained segmentation model, we could calculate the area of the wound. As to automatic wound diagnosis and treatment suggestion, based on updated clinical guidelines for pressure injuries, we proposed a flow chart for the diagnosis and treatment. It mainly comprised two classification tasks: the erythema classification task and the necrotic tissue classification task. We trained and compared 4 classic architectures of CNN (Convolutional Neural Network): AlexNet, VGG, ResNet and Inception-ResNet-V2. We compared the performance of above tasks and models. On the segmentation task, U-Net had better performance than Mask R-CNN (IoU: 0.7773 versus IoU: 0.4604). For automatic wound area measurement, our estimation had a mean relative error (MRE) about 26.2%. On the erythema and the necrotic tissue classification tasks, Inception-ResNet-V2 had the best performance and got high accuracy about 98.5% and 97%, respectively. Due to above three successful studies and validations, we considered our Pressure-Injury Assessment System (PIAS) could give acceptable wound assessment and give treatment suggestions that are worthwhile of consideration. To determine whether our PIAS can be used in a clinical setting, whether it can assist first-line caregivers, and whether it can enhance overall treatment and care, more prospective trials are required.