基於三維卷積神經網路於電腦斷層影像腎臟腫瘤偵測與量化;Renal Tumor Detection and Quantification in Computed Tomography Based on Three-Dimensional Convolutional Neural Networks

NCUIR > College of Electrical Engineering & Computer Science > Graduate Institute of Electrical Engineering > Electronic Thesis & Dissertation > Item 987654321/93371

Please use this identifier to cite or link to this item: http://ir.lib.ncu.edu.tw/handle/987654321/93371

Title:	基於三維卷積神經網路於電腦斷層影像腎臟腫瘤偵測與量化;Renal Tumor Detection and Quantification in Computed Tomography Based on Three-Dimensional Convolutional Neural Networks
Authors:	魏銘皓;Wei, Ming-Hao
Contributors:	電機工程學系
Keywords:	腎臟腫瘤;電腦斷層;深度學習;自動分割;圖形使用者介面;Renal Tumor;Computed Tomography;Deep Learning;Automated Segmentation;Graphical User Interface
Date:	2023-07-28
Issue Date:	2024-09-19 16:56:35 (UTC+8)
Publisher:	國立中央大學
Abstract:	腎臟腫瘤(Renal tumor)為腎臟上病變而發展出的區域，此區域有分成良性惡性，若可以及早發現和治療即可減少許多未來的風險以及成本，那該如何及早發現和治療就需要想該如何準確和有效率的辦法來實現。在電腦斷層影像識別腎臟腫瘤相較於其他檢查方式而言，。本研究共採取 382 筆病患資料，先採取對電腦斷層影像上做前處理；接著利用三維卷積神經網路(3D Convolutional Neural Network, 3D CNN)架構的 Deepmedic 採兩階段式的分割。第一階段目的為提取包含腎臟與腫瘤的感興趣區域(Region of interest, ROI)，接著透過影像後處理排除體積較小的偽陽性區域；第二階段再以第一階段所提取的感興趣區域進一步去分割腎臟腫瘤。最後再將此自動分割腎臟腫瘤的功能以更直觀、更方便使用的想法上去建立一個圖形使用者介面，以利於醫師能直觀上能夠容易上手。本文數據使用患有腎臟腫瘤病人的電腦斷層影像，其中 60%為訓練集、 20%為驗證集以及 20%為測試集。腎臟和腫瘤的感興趣區域分割後經由後處理的效能評估平均戴斯係數為 93.82%、平均精確率為 94.86%以及平均召回率為 93.66%；腎臟腫瘤分割效能的效能評估平均戴斯係數為 88.19%、平均精確率為 90.36%以及平均召回率為 88.23%。透過深度學習的方式去訓練模型，使得可以在電腦斷層影像上分割腎臟腫瘤的位置並且量化體積，可以看到腎臟和腫瘤的區域分割效能不錯，能 ii 夠有效地抓出腎臟和腫瘤區域。最後再製作出一個能使醫師更簡單且直接操作的使用者介面，達到讓醫師能夠在短時間內有著極高的效率。本研究基於深度學習對電腦斷層影像進行腎臟腫瘤自動分割與量化，進而輔助醫師於臨床診斷上以及相關應用。 ;Renal tumor refers to a localized abnormality that develops in the kidney, which can be classified as either benign or malignant. Early detection and treatment of renal tumors are crucial to mitigate future risks and reduce associated costs. To achieve accurate and efficient early detection and treatment, computerized tomography (CT) imaging plays a significant role compared to other diagnostic methods. In this study, a total of 382 patient data were collected, and preprocessing was performed on the CT images. A two-stage segmentation approach utilizing a Three-Dimensional Convolutional Neural Network (3D CNN) architecture called Deepmedic was employed. The first stage aimed to extract the Region of Interest (ROI) containing the kidney and the tumor. Subsequently, post-processing was conducted to eliminate small false positive regions. In the second stage, the ROI obtained from the first stage was further segmented to delineate the renal tumor. Finally, an intuitive and user-friendly graphical user interface was developed to facilitate the ease of use for physicians. The data used in this study consisted of CT images from patients with renal tumors, with 60% used for training, 20% for validation, and 20% for test set. The performance evaluation of the segmented kidney and tumor regions yielded an average Dice coefficient of 93.82%, an average precision of 94.86%, and an average recall of 93.66%. For the renal tumor segmentation, the average Dice coefficient was 88.19%, the average precision was 90.36%, and the average recall was 88.23%. By employing deep learning, the model was trained to segment the location of renal tumors and quantify their volume on CT images, showing favorable iv segmentation performance for the kidney and tumor regions. The developed user interface further simplified the interaction for physicians, enabling them to achieve high efficiency within a short timeframe. This research contributes to the automated segmentation and quantification of renal tumors using deep learning on CT images, offering valuable support for clinical diagnosis and related applications.
Appears in Collections:	[Graduate Institute of Electrical Engineering] Electronic Thesis & Dissertation

Files in This Item:

File	Description	Size	Format
index.html		0Kb	HTML	19	View/Open

社群 sharing

Loading...