- 學位論文
一個基於深度學習的腸胃道影像疾病檢測方法
A Deep Learning Approach for Disease Detection with Gastrointestinal Images
摘要
膠囊內視鏡與傳統有線式內視鏡都是現今常用的腸胃道檢測方法，隨著光學鏡頭製程技術的進步，內視鏡檢測效能也迅速提升。而要精確的從腸胃道影像或影片中找出腸胃道疾病與問題至今仍是關鍵的課題，甚至是影響醫生最終的疾病判斷與治療用藥，但是由於一次檢測就會產生數萬張影像和長時間的檢測影片，逐張影像檢查或長時間注視著螢幕觀看，造成醫師診斷上很大的負擔，並且也不符合效益。因此，本研究旨在研發出內視鏡影像的自動辨識系統，減少醫師進行診察所需觀看的影像數量，藉此改善醫師在診斷過程花費大量時間的問題，爭取醫療的時效性。經過多次與醫生面談後得出結論，現在主流的疾病物件偵測並無法滿足醫生需求，更容易誤判造成醫生困擾，所提出的系統只需要判斷出是否有異常即可，而疾病判斷的責任還是交還給醫生。 本研究提出一套系統，利用深度學習網路來做內視鏡影像辨識。在此處是用卷積神經網路(CNN, Convolutional neural network)所衍伸出的各種不同網路模型如VGG16和VGG19。第一部份的研究，是利用同一筆腸胃道影像Kvasir資料庫，比較出何種網路更適合用在醫療用腸胃道影像，而第二部份研究則是關於內視鏡影像的正常與異常檢測，此部份才是真正能帶給醫生實際幫助的研究。 實驗結果顯示，在第一部份Kvasir資料庫的辨識中VGG16與VGG19皆表現出不錯的性能，而在第二部份正常與異常辨識則是VGG16有最好的表現。結果顯示VGG16搭配上來自於ImageNet的Transfer Learning最適合用在內視鏡影像的自動檢測上，正確率最高來到94.60%。
並列摘要
Both capsule endoscopes and traditional wired endoscopes are commonly used as gastrointestinal inspection methods today. With the advancement of optical lens manufacturing process technology, the inspection performance of endoscopes has also improved rapidly. It is still a key issue to accurately find gastrointestinal diseases and problems from gastrointestinal images or videos, and it even affects the doctor’s final disease judgment and treatment medication. However, a single test will produce tens of thousands of images and long time-testing videos, checking images one by one, or watching the screen for a long time, cause a great burden on the doctor's diagnosis, and it is also not effective. Therefore, this study aims to develop an automatic recognition system for endoscopic images to reduce the number of images that doctors need to watch for diagnosis, so as to improve the problem that doctors spend a lot of time in the diagnosis process and strive for the timeliness of medical treatment. After many interviews with doctors, it is concluded that the current mainstream methodology for disease object detection cannot meet the needs of doctors, and it is easier to misjudge and cause doctors’ troubles. The proposed system only needs to determine whether there is abnormality, and the responsibility for disease judgment is returned to the doctor. This study proposes a system that uses deep learning networks for endoscopic image recognition. Here we use various network models derived from convolutional neural network (CNN) such as VGG16 and VGG19. The first part of the study is to use the same Kvasir database of gastrointestinal images to determine which network is more suitable for medical gastrointestinal images, while the second part is about the normal and anomaly detection of endoscopic images and the second part is the study that can really help doctors. The experimental results show that in the first part of the Kvasir database identification, both VGG16 and VGG19 show good performance, and in the second part of the normal and abnormal identification, VGG16 has the best performance. The results show that the VGG16 paired with Transfer Learning from ImageNet is the most suitable model for automatic detection of endoscope images, with the highest accuracy reaching 94.60%.
參考文獻
延伸閱讀
- Phan, N. N. (2022). 深度學習於生醫影像及多體學數據之應用 [doctoral dissertation, National Taiwan University]. Airiti Library. https://doi.org/10.6342/NTU202200364
- Yao, W. H. (2017). 以深度學習方法探索預測結腸癌復發的預後標記 [master's thesis, National Tsing Hua University]. Airiti Library. https://www.airitilibrary.com/Article/Detail?DocID=U0016-0401201815533439
- Wu, S. Y. (2021). 以深度學習方法發展之電腦輔助瘜肉定位系統應用於大腸鏡切除手術教學 [master's thesis, National Taiwan University]. Airiti Library. https://doi.org/10.6342/NTU202102285
- 李育倫.（2022）。基於深度學習在內視鏡檢查中估計覆蓋率之研究〔碩士論文，國立臺灣大學〕。華藝線上圖書館。https://doi.org/10.6342/NTU202202979
- 蕭輔仁、廖珩君、廖俊智（2020）。Clinical Application of Deep Learning for Brain Image。台灣醫學，24(1)，50-57。https://doi.org/10.6320/FJM.202001_24(1).0005