臺灣博碩士論文加值系統

目的：在過往的研究中發現傷口與周遭組織之間的溫度是具有關聯的，但在臨床上多半採用醫師的體感比較傷口與周邊軀幹的溫度差異，傳統上為單純計算兩者的溫度差，並使用固定閾值進行傷口感染與組織缺血的分類，本論文目的在於導入AI輔助臨床醫師進行診斷與評估，讓模型透過學習紅外線熱顯像圖，自動找出傷口分類與溫度分布關聯，最終輸出傷口之分類以提供醫師進一步之處置。
材料與方法:本研究使用的資料集之機構審查委員會(IRB)編號為KMUHIRB-E(I)20220033，研究使用資料由高雄醫學大學附設中和紀念醫院提供，資料為臨床醫師透過安裝在手機上的熱顯像儀對臨床病患之患部與周遭組織進行拍攝，拍攝距離患者患部約45-60 cm，病患人次為82位，包含腳掌、小腿、手掌、前臂、膝蓋、大腿、陰部與頭部共104張影像。取得傷口組織的溫度分佈資訊並結合深度學習方法與數值分析技術，使用本論文提出之區域特徵萃取深度融合網路進行傷口特徵的框選、萃取與分類。
結果：本論文使用二階段檢測方法進行傷口的分類，第一階段興趣區域檢測mAp50為0.837±0.119，經由特徵萃取與深度融合網路進行訓練後接上強分類器進行特徵篩選與分類，其模型分類準確率為0.83±0.110，AUC為0.81，本論文的效果能夠有效提供臨床醫師進行初步判斷並選擇合適處置方式。

Objective: In previous studies, it was found that the temperature between the wound and the surrounding tissue is related. However, in clinical practice, the doctor's body sensation is mostly used to compare the temperature difference between the wound and the surrounding torso. Traditionally, the temperature difference between the two is simply calculated, and A fixed threshold is used to classify wound infection and tissue ischemia. The purpose of this paper is to introduce AI to assist clinicians in diagnosis and evaluation, so that the model can automatically find the correlation between wound classification and temperature distribution by learning infrared thermal imaging images, and finally output the wound classification to provide further treatment by physicians.
Materials and Methods: The Institutional Review Board (IRB) number of the dataset used in this study is KMUHIRB-E(I)20220033. The research data is provided by Chung Ho Memorial Hospital, Kaohsiung Medical University. The instrument takes pictures of the affected part and surrounding tissues of clinical patients. The shooting distance is about 45-60 cm from the affected part of the patient. The number of patients is 82, including 104 photos of the soles of the feet, calves, palms, forearms, knees, thighs, genitals and heads. image. The temperature distribution information of wound tissue is obtained and combined with deep learning method and numerical analysis technology, the deep fusion network of regional feature extraction proposed in this paper is used to select, extract and classify wound features.
Result: This paper uses a two-stage detection method to classify wounds. The mAp50 of the first-stage interest region detection is 0.837±0.119. After training through feature extraction and deep fusion network, it is connected to a strong classifier for feature screening and classification. The model classification is accurate The accuracy rate is 0.83±0.110, and the AUC is 0.81. The effect of this paper can effectively provide clinicians with preliminary judgment and selection of appropriate treatment methods.

目錄
中文摘要 I
Abstract II
誌謝 IV
目錄 V
圖目錄 IX
表目錄 XIII
名詞縮寫表 XV
第一章 緒論 1
1.1 研究背景與動機目的 1
1.2 研究流程 2
1.3 論文架構 3
第二章 文獻探討 4
2.1 紅外線熱顯像儀臨床應用 4
2.2 臨床病症判斷方式 6
2.3 特徵萃取器 8
2.3.1 Patch Embedding 8
2.3.2 ConvMixer 8
2.3.3 Dilated Convolution 11
2.3.4 CBAM 15
2.3.5 GELU 16
2.4 損失函數與優化器 18
2.4.1 FocalLoss 18
2.4.2 學習率優化 21
2.5 機器學習模型 24
2.5.1 XGBoost 24
2.5.2 CatBoost 25
2.6 ROI興趣域 26
2.6.1 Yolov5 26
2.6.2 YoloX 30
2.7 模型評估與驗證指標 32
2.7.1混淆矩陣 32
2.7.2 Specificity 33
2.7.3 Sensitivity 33
2.7.4 AUC ROC 34
第三章 研究架構與方法 35
3.1 數據蒐集 35
3.2 影像前處理 36
3.2.1 分離原始溫度資料 36
3.2.2 將溫度資訊上色 37
3.2.3 分離背景 38
3.2.4 使用數值分析輔助 41
3.2.5 遮罩修復 47
3.2.6 去背可視化效果 50
3.3 患部狀態分類 51
3.3.1 使用固定溫度閾值 51
3.3.2 使用常態分布檢視 54
3.4 ROI興趣域 55
3.4.1 框選出ROI興趣域 55
3.4.2 遷移式學習 56
3.4.3 框選座標映射 57
3.5 模型架構 59
3.5.1 訓練流程 59
3.5.2 ROI區域框選模組 60
3.5.3 特徵萃取模組 62
3.5.4 強分類器模組 68
3.6 提升方法與比較 69
3.6.1 遷移學習ROI框選效果提升 69
3.6.2 空洞卷積 73
3.6.3 自注意力機制 74
3.6.4 學習率調整 75
第四章 研究結果與討論 77
4.1 使用數值分式分析結果 77
4.2 深度學習分類結果 80
4.2.1 模型驗證方式 80
4.2.2 特徵萃取器基礎模型架構 81
4.2.3 色彩地圖映射比較 82
4.2.4 卷積核大小之視野域對模型比較 83
4.2.5 卷積核擴張率對模型比較 84
4.2.6 注意力機制對模型的提升 85
4.2.7 強分類器對模型的提升 86
4.2.8 模型注意力可視化效果 87
4.2.9 模型提升效果 89
第五章 結論與未來展望 90
5.1 結論 90
5.2 未來展望 91
參考文獻 92

[1] van Netten, Jaap J., et al. "Diagnostic values for skin temperature assessment to detect diabetes-related foot complications." Diabetes technology & therapeutics 16.11 (2014): 714-721.
[2] Stojcsics, Daniel, et al. "High resolution 3D thermal imaging using FLIR DUO R sensor." 2018 IEEE 22nd International Conference on Intelligent Engineering Systems (INES). IEEE, 2018.
[3] ZHU, WEI‐PING, and XIANG‐RONG XIN. "Study on the distribution pattern of skin temperature in normal Chinese and detection of the depth of early burn wound by infrared thermography." Annals of the New York Academy of Sciences 888.1 (1999): 300-313.
[4] M.E.H. Jaspers, et al. 2017, “The FLIR ONE thermal imager for the assessment of burn wounds: Reliability and validity study” Journal of the International Society for Burn Injuries(Burns) 43, 2017, 1516 -1523.
[5] Dosovitskiy, A., et al. 2020. “An Image is Worth 16x16 Words: Transformers for Image Recognition at Scale” International Conference on Learning Representations (ICLR 2021).
[6] Shancheng Fang, et al. 2021 “Read Like Humans: Autonomous, Bidirectional and Iterative Language Modeling for Scene Text Recognition” IEEE Conference on Computer Vision and Pattern Recognition(CVPR 2021).
[7] Alex Krizhevsky, et al. 2012.” ImageNet Classification with Deep Convolutional Neural Networks”, Advances in Neural Information Processing Systems 25 (NIPS 2012).
[8] Kaiming He, et al. 2015.” Deep Residual Learning for Image Recognition” IEEE Conference on Computer Vision and Pattern Recognition(CVPR 2016).
[9] Fisher Yu, et al. 2016.”MULTI-SCALE CONTEXT AGGREGATION BY DILATED CONVOLUTIONS”, International Conference on Learning Representations (ICLR 2016).
[10] Panqu Wang, et al. 2018, ” Understanding Convolution for Semantic Segmentation” IEEE Winter Conference on Applications of Computer Vision (WACV 2018).
[11] Sanghyun Woo, et al. 2018,” CBAM: Convolutional Block Attention Module”, Europeon Conference on Computer Vision (ECCV 2018)
[12] Jie Hu, et al. 2017, “Squeeze-and-Excitation Networks” IEEE Conference on Computer Vision and Pattern Recognition(CVPR 2018).
[13] Srivastava, Nitish, et al. "Dropout: a simple way to prevent neural networks from overfitting." The journal of machine learning research 15.1 (2014): 1929-1958.
[14] Hendrycks, Dan, and Kevin Gimpel. "Gaussian error linear units (gelus)." arXiv preprint arXiv:1606.08415 (2016).
[15] Lin, Tsung-Yi, et al. "Focal loss for dense object detection." Proceedings of the IEEE international conference on computer vision. (ICCV 2017).
[16] Loshchilov, Ilya, et al. "Decoupled weight decay regularization”, International Conference on Learning Representations (ICLR 2019).
[17] Goyal, Priya, et al. "Accurate, large minibatch sgd: Training imagenet in 1 hour." arXiv preprint arXiv:1706.02677 (2017).
[18] Prokhorenkova, Liudmila, et al. "CatBoost: unbiased boosting with categorical features." Advances in neural information processing systems 31 (2018).
[19] Lin, Tsung-Yi, et al. "Feature pyramid networks for object detection." Proceedings of the IEEE conference on computer vision and pattern recognition. (CVPR 2017).
[20] Ren, Shaoqing, et al. "Faster r-cnn: Towards real-time object detection with region proposal networks." Advances in neural information processing systems 28 (2015).
[21] Rezatofighi, Hamid, et al. "Generalized intersection over union: A metric and a loss for bounding box regression." Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. (CVPR 2019).
[22] Ge, Zheng, et al. "Yolox: Exceeding yolo series in 2021." arXiv preprint arXiv:2107.08430 (2021).
[23] Cao, Zhe, et al. "Realtime multi-person 2d pose estimation using part affinity fields." Proceedings of the IEEE conference on computer vision and pattern recognition (CVPR 2017).
[24] XuebinQin, et al. “U2-Net: Going deeper with nested U-structure for salient object detection”. Pattern Recognition, Volume 106, October 2020, 107404
[25] H.C Shin, et al. 2016, “Deep Convolutional Neural Networks for Computer-Aided Detection: CNN Architectures, Dataset Characteristics and Transfer Learning”, IEEE Transactions on Medical Imaging, vol. 35, no. 5, pp. 1285-1298, May 2016, doi: 10.1109/TMI.2016.2528162.
[26] Tzeng, Eric, et al. "Deep domain confusion: Maximizing for domain invariance." arXiv preprint arXiv:1412.3474 (2014).
[27] Xue, Erica Y., et al. “Use of FLIR ONE smartphone thermography in burn wound assessment.” Annals of plastic surgery 80.4 (2018): S236-S238.
Bochkovskiy, Alexey, Chien-Yao Wang, and Hong-Yuan Mark Liao. "Yolov4: Optimal speed and accuracy of object detection." arXiv preprint arXiv:2004.10934 (2020).