- 學位論文
深度學習應用於燒傷傷口影像辨識系統
Burn Wound Image Detection System Using Deep Learning
摘要
在處理燒傷傷口時，醫護人員對於傷口處理與面積有不同之定義與判斷，加上非專業醫師之判斷，會提高傷口治療之成本。因此，在醫療過程中可以透過深度學習之應用來協助醫護人員來判斷患者之燒傷傷口面積，為減少額外的成本。 本論文研究主要目標為了透過深度學習之模型，打造了燒傷傷口範圍之辨識系統來辨識傷口區域。我們也有透過COCO與 ImageNet之類神經權重做遷移學習的部分。在資料訓練中也透過各種不同之資料曾強化 (Data Augmentation) 提高訓練效果，讓模型學習在實際拍攝影像之不同情況。我們對於燒傷傷口感興趣之範圍(Region of Interest)之研究有試著兩種不同之深度學習架構 : (1)實例分割遮罩型區域類神經網路(Mask Region-based Convolution Neural Network, Mask-RCNN)，(2) 語意分割遮罩型區域類神經(U-net)。且透過辨識患者之手掌面積做比較，可判斷患者在身上之燒傷傷口在全身佔的百分比比例為多少。我們在系統架構上之模型部署都統一使用了Mask RCNN之模型為辨識燒傷與手掌之區域。
並列摘要
In the case of treating burn wounds, medical staffs have different definitions and judgments regarding wound treatment and its area, adding to the judgment of the non-professional doctors, the cost of the wound treatment will increase. Therefore, we can implement a state-of-art deep learning model to assist medical staff in judging the burn area of the patient in order to reduce additional costs. In this study, our main goal is to implement a deep learning model to create a burn wound area detection system to detect the burned part of the images. In the process of data collecting, we also have met the dilemma of difficult characteristics of the burn itself and different numbers of images for each class of data. And in order to improve training efficiency, we have implemented transfer learning of COCO and ImageNet neural network initial weight to our training model. We have implemented two types of Region of Interest (ROI) deep learning model, the first is Mask-RCNN (Mask Region-based Convolution Neural Network), the second is U-Net (U-shaped Network). We also implement a model to detect the hand palm of the patient, to comparing the area of the hand palm of the patient’s hand towards their burn areas, then we can determine the percentage of the patient’s burn wounds on their body.