Please use this identifier to cite or link to this item:
http://tdr.lib.ntu.edu.tw/jspui/handle/123456789/84290
Title: | 應用多群聚合及空間金字塔池化之3-D ResNeSt於自動乳房超音波電腦輔助腫瘤診斷系統 Computer-aided Tumor Diagnosis Using 3-D ResNeSt with Multi-group Aggregation and Spatial Pyramid Pooling for Automated Breast Ultrasound |
Authors: | 戴勤 Chin Dai |
Advisor: | 張瑞峰 Ruey-Feng Chang |
Keyword: | 乳癌,自動乳房超音波,電腦輔助診斷系統,卷積神經網路, Breast cancer,Automated breast ultrasound (ABUS),Computer-aided diagnosis (CADx),Convolution neural network (CNN), |
Publication Year : | 2021 |
Degree: | 碩士 |
Abstract: | 乳癌是女性最常見的疾病之一,透過早期的檢查與治療,可以有效降低死亡率。而全自動乳房超音波系統(Automated Breast Ultrasound System, ABUS)是目前應用最廣泛的乳癌檢查方法之一。然而,當醫生透過超音波影像進行診斷時,容易受到腫瘤的形狀和紋理影響,導致影像判讀有極大的差異。為了解決這個問題,電腦輔助診斷系統(Computer-Aided Diagnosis system)被開發出來以輔助醫生做診斷。近年來,基於卷積神經網路(Convolutional Neural Network, CNN)的電腦輔助診斷系統成為趨勢,並成功地應用於醫學影像。因此,本研究提出一個基於神經網路的電腦輔助診斷系統協助乳癌診斷。
我們的電腦輔助診斷系統包含影像前處理、三維腫瘤切割網路和三維腫瘤分類網路。在影像前處理中,首先會先由經驗豐富的醫生提取腫瘤區域,而後會將腫瘤區域影像固定至相同大小,並利用直方圖均衡化處理影像。然後將調整後的影像送到分割模型中得到相對應的腫瘤遮罩。最後將調整大小後的原始影像、均衡化處理的影像以及腫瘤遮罩輸入到我們的分類模型中,以判斷腫瘤為良性或惡性。實驗結果顯示,我們所提出的電腦輔助診斷系統準確率為89.9%、靈敏性為88.9%、特異性為90.4%。結果表明,此系統有良好的性能且可做為臨床醫生決策參考。 Breast cancer is one of the most common diseases in women. Through the early examination and treatment, the mortality could be effectively reduced. The automated breast ultrasound (ABUS) was one of the most widely-used examinations for breast cancer. However, physicians were susceptible to the shape and texture of tumors and led to large variations in image interpretation. The computer-aided diagnosis (CADx) system was developed to offer a second opinion to deal with this problem. In recent years, applying convolution neural network (CNN) on CADx system became a tendency and achieved great success on medical images. Hence, a CNN-based CADx system was proposed for breast tumor diagnosis in this study. Our CADx system was a 3-D model and consisted of data preprocessing, 3-D tumor segmentation, and 3-D tumor classification. In the data preprocessing, the volume of interest (VOI) was firstly extracted by the experienced physicians, and the tumor VOI would be resized and conducted histogram equalization. After that, the resized VOI would be sent into the segmentation model to obtain the corresponding tumor mask. Lastly, the resized VOI, the equalized VOI, and the tumor mask were fed into our classification model, 3-D MASP-ResNeSt, to determine the tumor was benign or malignant. In our experiment, the proposed CADx system achieved 89.6% accuracy, 88.9% sensitivity, and 90.4% specificity. The results indicated the proposed CADx system had a good performance and might be a second opinion for physicians to make the decision. |
URI: | http://tdr.lib.ntu.edu.tw/jspui/handle/123456789/84290 |
DOI: | 10.6342/NTU202200941 |
Fulltext Rights: | 同意授權(限校園內公開) |
metadata.dc.date.embargo-lift: | 2024-06-16 |
Appears in Collections: | 生醫電子與資訊學研究所 |
Files in This Item:
File | Size | Format | |
---|---|---|---|
ntu-110-2.pdf Access limited in NTU ip range | 1.25 MB | Adobe PDF | View/Open |
Items in DSpace are protected by copyright, with all rights reserved, unless otherwise indicated.