運用影像組學與機器學習於聽神經瘤放射手術預後與腦部海綿狀血管瘤癲癇風險之研究;Radiomics and Machine Learning Approaches for Predicting Radiosurgical Prognosis in Vestibular Schwannoma and Epileptogenic Risk in Cerebral Cavernous Malformations

NCU Institutional Repository > 資訊電機學院 > 電機工程研究所 > 博碩士論文 > Item 987654321/98555

jsp.display-item.identifier=請使用永久網址來引用或連結此文件: https://ir.lib.ncu.edu.tw/handle/987654321/98555

题名:	運用影像組學與機器學習於聽神經瘤放射手術預後與腦部海綿狀血管瘤癲癇風險之研究;Radiomics and Machine Learning Approaches for Predicting Radiosurgical Prognosis in Vestibular Schwannoma and Epileptogenic Risk in Cerebral Cavernous Malformations
作者:	吳明宏;Wu, Ming-Hong
贡献者:	電機工程學系
关键词:	磁振造影;聽神經瘤;腦部海綿狀血管瘤;機器學習;Magnetic Resonance Imaging;Vestibular Schwannoma;Cerebral Cavernous Malformation;Machine Learning
日期:	2025-08-22
上传时间:	2025-10-17 12:55:30 (UTC+8)
出版者:	國立中央大學
摘要:	放射手術為治療聽神經瘤與腦部海綿狀血管瘤常用之非侵入性療法，其具有高局部控制率與低手術風險等優勢，然而實務上術後預後反應差異性仍大，尤其對於腫瘤持續增長或癲癇發作的風險評估，主要仰賴臨床經驗與影像觀察，缺乏客觀、量化且可自動化之輔助工具，限制了個別化治療策略的制定。本研究擷取患者接受加馬刀放射手術前的磁振造影影像，包含增強型T1加權影像（contrast-enhanced T1-weighted imaging, CET1WI）與T2加權影像(T2-weighted imaging, T2WI)，進行影像預處理（對位、Z-score標準化、空間標準化）後，提取形狀、強度與紋理等影像組學特徵，並結合機器學習模型建立預測模型。同時針對高度不平衡的二元分類問題，導入合成少數類別過採樣技術(Synthetic Minority Over-sampling Technique, SMOTE)、邊界合成少數類別過採樣技術(Borderline-SMOTE)、安全層級合成少數類別過採樣技術(Safe-Level-SMOTE)、自適應合成採樣技術(Adaptive Synthetic Sampling, ADASYN)等過採樣方法對特徵空間生成少數類別新的資料樣本，並結合卡方檢定(Chi-square, Chi2)、最小冗餘最大相關性(minimum Redundancy Maximum Relevance, mRMR)、鄰近成分分析(Neighborhood Component Analysis, NCA)等特徵選擇策略以挑選最具預測力的特徵。模型效能透過留一交叉驗證進行評估，以確保在小樣本情境下的穩健性。在聽神經瘤研究中，主要預測放射手術後腫瘤是否持續生長，最佳模型為結合ADASYN資料增量、Chi-square特徵篩選與Linear SVM之架構，其在預測術後是否持續生長方面表現優異，達成準確率0.869、精確率0.512、靈敏性0.933、特異性0.859與 AUC 值為 0.897，顯示模型具有良好判別能力，對於術後高風險個案具備鑑別潛力。在腦部海綿狀血管瘤研究中，主要預測是否可能癲癇，本研究進一步納入腫瘤位置與灰質排擠體積等空間特徵，提升模型對腦部區域侵犯程度的敏感性。最佳模型同樣採用Linear SVM，並搭配Safe-Level-SMOTE與Chi-square特徵選擇，於預測是否發生癲癇反應中，達到準確率0.837、精確率0.409、靈敏性1.000 、特異性0.816與 AUC 0.974，展現高度分類能力與臨床應用潛力。綜合而言，本研究提出一套結合MRI影像組學與機器學習之預測流程，針對聽神經瘤與腦部海綿狀血管瘤患者於放射手術後的反應進行量化預測。透過有效的特徵選擇與資料增量技術，所建構之模型展現良好準確率與靈敏性，顯示本方法有助於彌補傳統臨床經驗評估之主觀限制，提供一個具客觀性、自動化且具臨床應用潛力的輔助工具，進一步推動個別化治療策略之實現。 ;Radiosurgery is a commonly used non-invasive treatment for vestibular schwannomas and cerebral cavernous malformations, offering high local control rates and low surgical risks. However, in clinical practice, postoperative outcomes remain highly variable. In particular, risk assessment for tumor progression or seizure occurrence largely relies on clinical experience and visual interpretation of imaging, lacking objective, quantitative, and automated tools to support personalized treatment planning. In this study, preoperative magnetic resonance imaging (MRI) scans—including contrast-enhanced T1-weighted and T2-weighted images—were collected from patients undergoing Gamma Knife radiosurgery. Following image preprocessing (registration, Z-score normalization, and spatial normalization), radiomic features such as shape, intensity, and texture were extracted. These features were then used to develop predictive models through machine learning approaches. To address the highly imbalanced nature of the binary classification problem, several oversampling techniques were applied to generate synthetic minority class samples in the feature space. These included the Synthetic Minority Over-sampling Technique (SMOTE), Borderline-SMOTE, Safe-Level-SMOTE, and Adaptive Synthetic Sampling (ADASYN). Feature selection strategies such as Chi-square test (Chi2), minimum Redundancy Maximum Relevance (mRMR), and Neighborhood Component Analysis (NCA) were employed to identify the most predictive features. Model performance was evaluated using leave-one-out cross-validation (LOOCV) to ensure robustness in the context of limited sample sizes. In the vestibular schwannoma study, the primary objective was to predict post-radiosurgery tumor progression. The best-performing model combined ADASYN for data augmentation, Chi-square for feature selection, and a Linear Support Vector Machine (SVM). This model demonstrated strong predictive performance with an accuracy of 0.869, precision of 0.512, sensitivity of 0.933, specificity of 0.859, and an AUC of 0.897—indicating excellent discriminative ability and potential for identifying high-risk patients post-treatment. In the cerebral cavernous malformation study, the goal was to predict seizure risk. Spatial features such as tumor location and gray matter displacement volume were further incorporated to enhance the model’s sensitivity to regional brain involvement. The best-performing model also utilized a Linear SVM, combined with Safe-Level-SMOTE and Chi-square feature selection. It achieved an accuracy of 0.837, precision of 0.409, sensitivity of 1.000, specificity of 0.816, and an AUC of 0.974—demonstrating high classification capability and strong potential for clinical application. In summary, this study proposes a predictive framework that integrates MRI-based radiomics with machine learning to quantitatively assess post-radiosurgery outcomes in patients with vestibular schwannoma and cerebral cavernous malformation. By incorporating effective feature selection and data augmentation techniques, the constructed models demonstrated high accuracy and sensitivity. These findings suggest that the proposed approach can help overcome the subjectivity of traditional clinical assessments, offering an objective, automated, and clinically applicable decision-support tool to facilitate the implementation of personalized treatment strategies.
显示于类别:	[電機工程研究所] 博碩士論文

文件中的档案:

档案	描述	大小	格式	浏览次数
index.html		0Kb	HTML	10	检视/开启

在NCUIR中所有的数据项都受到原著作权保护.

社群 sharing

数据加载中.....