應用深度學習於傷口影像自動判讀

Abstract

在傷口護理中，無論是慢性傷口還是手術傷口，皆需要追蹤傷口癒合情況並給予相應治療，患者還需定期到醫院詢問醫師的建議。因此，醫療中心需要一個傷口護理管理系統來打破醫療服務中時間和地理的限制，為患者提供更好的醫療服務。 在這項研究中，我們提出了一個基於深度全卷積網絡（Fully Convolutional Network; FCN）的手術傷口影像自動判讀系統。由於傷口數據集的資料數量有限，我們應用遷移學習(Transfer Learning)先訓練神經網路於PASCAL VOC 2012，然後再微調訓練我們的神經網絡。我們的系統可以檢測皮膚上的異常感興趣區域（Region of Interest; ROI）並協助傷口偵測和感染判讀。在傷口偵測中，我們可以擷取出準確的手術傷口區域並預測傷口內縫合處的實際長度（毫米）。在感染判讀中，我們定義了9種感染類型，並使用基於區域的分類器來偵測傷口和周圍皮膚的感染區域。為了收集判讀模型的訓練資料，我們與專業臨床醫師合作並開發了傷口資料收集系統。我們的模型使用5次交叉驗證，異常區域偵測達到0.775（±0.02）的平均覆蓋率(mean-IoU)，並且在皮膚上的手術傷口與感染區域有0.84（±0.03）的涵蓋率(coverage)。在傷口區域檢測中我們達成了0.609（±0.017）平均IoU並且於長度預測的平均誤差只有8.59（±1.151）毫米。在感染判讀中，我們關注在三種主要感染類型，並達到0.671（±0.017）平均覆蓋率。

In wound care, whether it is chronic wound or surgical wound, the principle of caregiving is to keep track of the state of wound healing and give the corresponding treatment, the patients also need to visit hospital regularly for the suggestion from clinicians. Therefore, medical centers need a wound care management system to break the restrictions of time and geography in medical service, providing a better medical care for patients. In this study, we propose an automatic interpretation system for surgical wound image based on deep Fully Convolutional Network (FCN). Because of the limit wound dataset, we apply the Transfer Learning to use the pre-trained FCN on PASCAL VOC 2012, then fine-tune the network in our task. Our system can detect abnormal ROI (Region of Interest) on the skin and assist in wound detection and infection interpretation. In wound detection, we can extract the accurate region of surgical wound and predict the actual length (mm) of seam within the wound. In infection interpretation, we defined 9 types of infection and used the region-based classification to detect the infection ROIs in the wound and surrounding skin. To collect the training data for interpretation models, we cooperate with expert clinicians and develop a Wound Data Collection System. Our models use the 5-fold cross-validation, the abnormal ROI detection achieves 0.775 (± 0.02) mean-IoU with 0.84 (± 0.03) coverage of surgical wound and infection ROI on the skin. The wound ROI detection achieves the 0.609 (± 0.017) mean-IoU and the mean error 8.59 (± 1.151) mm. In infection interpretation, we focus on the three major infection types, and achieve the 0.671 (± 0.017) mean-IoU.