以三維卷積神經網路探究大腦微出血之偵測與量化;Cerebral Microbleed Detection and Quantification Based on Three-Dimensional Convolutional Neural Networks

NCUIR > College of Electrical Engineering & Computer Science > Graduate Institute of Electrical Engineering > Electronic Thesis & Dissertation > Item 987654321/89948

Please use this identifier to cite or link to this item: http://ir.lib.ncu.edu.tw/handle/987654321/89948

Title:	以三維卷積神經網路探究大腦微出血之偵測與量化;Cerebral Microbleed Detection and Quantification Based on Three-Dimensional Convolutional Neural Networks
Authors:	黃裕凱;Huang, Yu-Kai
Contributors:	電機工程學系
Keywords:	大腦微出血;深度學習;人工智慧;自動偵測;圖形使用者介面;Cerebral microbleeds;deep learning;artificial intelligence;automatic detection;graphical user interface
Date:	2022-08-04
Issue Date:	2022-10-04 12:05:35 (UTC+8)
Publisher:	國立中央大學
Abstract:	大腦微出血(Cerebral Microbleeds, CMBs)為小型的慢性腦出血，近期研究中，已逐漸確認其與腦中風、智能障礙等因素有顯著的關係，及早檢測並進行治療成為一個日漸重要的課題。大腦微出血在敏感性加權血管成像(Susceptibility-weighted images, SWI)上表現最為明顯，一般呈現黑色均質圓形。本研究的目的是提出以SWI影像進行分析及自動偵測大腦微出血的方法，以人工智慧模型協助臨床醫師的辨識。本研究提出的方法主要分為三階段，第一階段為使用運行速度較快並可直接分割的Mask R-CNN，先對SWI影像進行實質腦提取，第二階段在實質腦區域範圍內利用YOLO架構去初步篩選疑似CMBs的部分，接著第三階段以三維卷積神經網路(3D Convolutional Neural Networks)進行最後的CMBs分類，從眾多疑似的CMBs中確認出真正的CMBs。本研究最後將這三階段的功能以圖形使用者介面方式呈現，方便使用者快速操作與觀看結果。本研究結果顯示SWI影像實質腦提取部份的精準度達98%，敏感度達93%，具95%的Dice係數；以YOLO網路於候選CMBs中篩選CMBs偵測其敏感度為90%，每位病人的平均偽陽數為78.19顆；最終3D CNN模型顯示大部分病人的CMB與非CMB都可被辨識分類，其敏感度為85%，每位病人平均偽陽數為2.41顆CMBs。總而言之，本研究利用神經網路進行自動檢測與標記，在大腦微出血的偵測獲得良好的結果，能將大部分的大腦微出血篩選出來，並排除疑似區域。在實質腦提取方面亦達到相當高的成果，能有效地將實質腦給提取出來。本研究將上述結果整合起來成功開發圖形使用者介面讓使用者可以簡單操作。 ;Cerebral Microbleed (CMB) is a small chronic cerebral hemorrhage. Recent studies have gradually confirmed that it has a significant relationship with factors such as stroke and intellectual disability. Early detection and treatment have become an increasingly important issue. Cerebral microbleeds are most evident on Susceptibility-Weighted Images (SWI) and usually appear as black homogeneous circles. The purpose of this study is to develop a method for automatically detecting cerebral microbleeds by analyzing SWI images and assisting clinicians to identify them with an artificial intelligence model. The method proposed in this study is mainly divided into three stages. The first stage is to use Mask R-CNN, which runs faster and can perform segmentation directly, to extract the parenchymal brain from the SWI image. The second stage uses YOLO to screen for candidate CMBs within the scope of the parenchymal brain area. The third stage uses a 3D convolutional neural network to classify the candidate CMBs to find the real CMBs. Finally, the functions of the three stages are integrated and presented through a graphical user interface to provide convenience and efficiency for the users. The results of the study showed that the accuracy of SWI image parenchymal brain extraction was 98%, the sensitivity was 93%, and the Dice coefficient was 95%; the sensitivity of detecting candidate CMBs on the YOLO network was 90%, and the average number of false positives per patient was 78.19 CMBs. The final 3D CNN model showed that CMB and non-CMB in most patients could be identified and classified, with a sensitivity of 85% and an average false positive average of 3.73 CMBs per patient. In summary, this research used three types of neural networks in series for automatic detection of CMBs. This method can largely reduce the burden of the clinicians in detecting CMBs. The output quality rendered by this method is good and clinically useful.
Appears in Collections:	[Graduate Institute of Electrical Engineering] Electronic Thesis & Dissertation

Files in This Item:

File	Description	Size	Format
index.html		0Kb	HTML	63	View/Open

社群 sharing

Loading...