http://tdr.lib.ntu.edu.tw/jspui/handle/123456789/89151 2026-04-23T01:25:21Z http://tdr.lib.ntu.edu.tw/jspui/handle/123456789/99927 標題: 開發一創新多模態深度學習方法提升卵巢腫瘤診斷精準度; Development of A Novel Multi-Modal Deep Learning Approach to Improve Diagnostic Precision in Ovarian Cancer 作者: 邱柏鈞; Po-Chun Chiu 摘要: 背景與目的：卵巢癌是台灣女性生殖器官惡性腫瘤的主要死因之一，早期診斷是對於改善患者預後的重要議題。然而，良性與惡性卵巢腫瘤的治療策略差異極大，術前診斷的準確性直接影響治療決策。傳統超音波診斷高度依賴操作者經驗，存在主觀性與變異性問題。本研究目的為開發一種結合超音波影像與臨床文字報告的多模態深度學習模型，以提升卵巢腫瘤良惡性診斷的精準度。 方法：本論文採用回溯性單中心設計，收集2011年至2021年間台大醫院系統中1,062名接受卵巢腫瘤手術患者的1,342張超音波影像及其相應的超音波文字報告。研究對象依據病理結果分類為良性組（n=612）及惡性組（n=450）後，建構多模態深度學習架構，同時整合DenseNet-121和Swin Transformer用於影像特徵提取，最終採用Bio-Clinical BERT處理臨床文字報告。模型採用受試者層級分層分割，並使用五倍交叉驗證，利用獨立測試集進行預測模型效果評估。 結果：本論文建構之多模態模型在受試者層級分類中達到81.68%準確率、79.38%敏感度、83.81%特異度及0.89的曲線下面積（AUC）。在影像層級分類中表現更佳，準確率為85.15%、敏感度86.73%、特異度83.65%、AUC為0.91。相較於單純影像模型，加入臨床文字資訊能顯著提升診斷效能。在變數重要性上，後向選擇分析顯示在文字報告上的子宮相關描述與卵巢腫瘤特徵描述對最終診斷皆具重要貢獻價值。 結論：本論文成功開發出一套創新整合超音波影像與臨床文字報告的多模態深度學習模型，其診斷效能優於傳統單一模態方法及現有臨床指引。此模型展現了人工智慧輔助診斷在卵巢腫瘤診斷上的應用潛力，有望成為臨床醫師重要的輔助工具，提升術前診斷精準度以改善患者照護品質。; Background and Objectives: In Taiwan, ovarian cancer represents the primary cause of mortality from gynecological malignancies among women, making early detection essential for enhancing patient prognosis. However, treatment strategies for benign versus malignant ovarian tumors differ significantly, making accurate preoperative diagnosis essential for appropriate clinical decision-making. Traditional ultrasound diagnosis is highly operator-dependent, introducing subjectivity and variability. To improve diagnostic precision in ovarian tumor classification, this research focuses on developing a multimodal deep learning system that combines ultrasound images with corresponding clinical text reports. Methods: Our study retrospectively analyzed 1,342 ultrasound images from 1,062 patients who received surgical treatment for ovarian tumors at the National Taiwan University Hospital system during the period from 2011 to 2021. Patients were classified into benign (endometrioma, n=612) and malignant (including borderline, n=450) groups based on pathological confirmation. A multimodal deep learning architecture was developed, incorporating DenseNet-121 and Swin Transformer for image feature extraction and Bio-Clinical BERT for processing clinical text reports. The dataset was split using subject-level stratification with 5-fold cross-validation and a 15% independent test set. Results: The multimodal model achieved superior performance at the subject level with 81.68% accuracy, 79.38% sensitivity, 83.81% specificity, and an area under the curve (AUC) of 0.89. Image-level classification demonstrated even better performance with 85.15% accuracy, 86.73% sensitivity, 83.65% specificity, and an AUC of 0.91. The integration of clinical text information significantly improved diagnostic performance compared to image-only models. Backward selection analysis revealed that both uterine findings and ovarian tumor descriptions contributed synergistically to the final diagnosis. Conclusions: This study successfully developed a multimodal deep learning model that integrates ultrasound images with clinical text reports, demonstrating superior diagnostic performance compared to traditional unimodal approaches and existing clinical guidelines. The model shows promising potential as an AI-assisted diagnostic tool for ovarian tumor classification, offering clinicians a valuable adjunctive tool to improve preoperative diagnostic accuracy and enhance patient care quality. 2025-01-01T00:00:00Z http://tdr.lib.ntu.edu.tw/jspui/handle/123456789/102104 標題: 遞迴雙變量 Probit 模型在二元工具變項統合分析之應用：以硬脊膜外麻醉對剖腹產之影響為例; Application of the Recursive Bivariate Probit Model in Instrumental Variable Meta-analysis for Binary Variables: A Case Study on the Effect of Epidural Analgesia on Cesarean Section 作者: 黃筱云; Siao-Yun Huang 摘要: 在醫學與流行病學的因果推論中，隨機對照試驗雖被視為黃金標準，但實務上常受限於受試者不遵循醫囑及倫理限制，導致傳統意向性治療分析（ITT）難以精確評估實際治療的生物機制。工具變數（Instrumental Variable, IV）方法雖可利用隨機分派作為工具來解決內生性問題，但現有的工具變數統合分析（Meta-IV）文獻多聚焦於連續型變數，缺乏針對工具變數、暴露與結果變數皆為二元（Binary Z-X-Y）且僅能取得匯總資料（Aggregate Data, AD）情境下的系統性分析框架。本研究旨在填補此方法學缺口，探討在該架構下如何建構適當的統計模型以維持估計的一致性。 本研究採用遞迴雙變量Probit模型（Recursive Bivariate Probit Model, RBPM）作為核心估計方法，該模型允許暴露與結果方程式的誤差項存在相關性，進而校正由未測量干擾因子所導致的內生性偏誤。研究程序包含模擬研究與實證資料分析兩部分。首先，透過蒙地卡羅模擬（Monte Carlo Simulation）在不同樣本數（N=1000, 3000, 5000）下比較 RBPM 與傳統Probit 模型的表現；其次，利用Zhou等人（2021）關於「硬脊膜外麻醉是否增加剖腹產風險」的臨床試驗匯總資料，透過2×2×2列聯表結構將其還原為個體層級資料（IPD），並應用RBPM檢驗內生性及估計真實因果效應。 模擬結果顯示，在存在未觀測干擾因子的情境下，忽略內生性的傳統模型會產生嚴重的高估偏誤，而RBPM則能有效還原真實的因果效應，且隨著樣本數增加，其估計精確度顯著提升。在實證分析方面，RBPM檢定結果顯示暴露與結果模型的誤差項存在顯著正相關（ρ=0.060），證實了內生性的存在，意即傾向選擇麻醉的產婦本身即具有較高的剖腹產潛在風險。然而，在校正此內生性偏誤後，研究發現實際接受硬脊膜外麻醉對於剖腹產的發生率並無統計上顯著的因果影響（p = 0.126）。 本研究結論指出，過去部分文獻認為麻醉會增加剖腹產風險的結論，可能源於未能妥善處理內生性所導致的選擇性偏誤。本研究透過RBPM模型證實了在二元變數架構下檢驗與校正內生性的必要性，並釐清了臨床上的爭議。此方法學嘗試展示了在特定條件下從匯總資料還原個體資料的可行性，為未來發展適用於更廣泛變數型態的工具變數統合分析統一框架提供了重要的實證基礎。; While Randomized Controlled Trials (RCTs) are regarded as the gold standard for causal inference in medicine and epidemiology, practical constraints such as subject noncompliance and ethical limitations often hinder traditional Intention-to-Treat (ITT) analysis from accurately evaluating the biological mechanisms of the actual treatment. Although Instrumental Variable (IV) methods can utilize random assignment as an instrument to address endogeneity, existing literature on IV meta-analysis (Meta-IV) predominantly focuses on continuous variables. There is a notable lack of a systematic analytical framework for scenarios where the instrument, exposure, and outcome are all binary (Binary Z-X-Y) and only Aggregate Data (AD) is available. This study aims to bridge this methodological gap by exploring the construction of appropriate statistical models within this framework to ensure consistent estimation. The Recursive Bivariate Probit Model (RBPM) was adopted as the core estimation method. This model allows for correlation between the error terms of the exposure and outcome equations, thereby correcting for endogeneity bias arising from unmeasured confounders. The research procedure comprised a simulation study and an empirical analysis. First, Monte Carlo simulations were conducted to compare the performance of RBPM with traditional Probit models across varying sample sizes (N = 1000, 3000, 5000). Second, utilizing aggregate clinical trial data from Zhou et al. (2021) regarding whether epidural analgesia increases the risk of cesarean section, the data was reconstructed into Individual Participant Data (IPD) via a 2 × 2 × 2 contingency table structure. RBPM was then applied to test for endogeneity and estimate the true causal effect. Simulation results demonstrated that in the presence of unobserved confounders, traditional models that ignore endogeneity yield severe overestimation bias. In contrast, RBPM effectively recovered the true causal effect, with estimation precision improving significantly as sample size increased. In the empirical analysis, RBPM results indicated a significant positive correlation (ρ = 0.060) between the error terms of the exposure and outcome models, confirming the presence of endogeneity. This implies that parturients who are inclined to choose epidural analgesia inherently possess a higher latent risk of cesarean section. However, after correcting for this endogeneity bias, the study found no statistically significant causal effect of receiving epidural analgesia on the incidence of cesarean sections (p = 0.126). The study concludes that previous findings suggesting anesthesia increases the risk of cesarean section may stem from selection bias resulting from a failure to adequately address endogeneity. By utilizing the RBPM, this study confirmed the necessity of testing for and correcting endogeneity within a binary variable framework, thereby clarifying clinical controversies. This methodological endeavor demonstrates the feasibility of reconstructing individual-level data from aggregate data under specific conditions, providing a substantial empirical foundation for the future development of a unified framework for instrumental variable meta-analysis applicable to a broader range of variable types. 2026-01-01T00:00:00Z http://tdr.lib.ntu.edu.tw/jspui/handle/123456789/99928 標題: 評估癲癇多基因風險分數在台灣族群與臨床樣本中預測及風險分層之表現; Evaluating the Predictive and Stratification Performance of Epilepsy Polygenic Risk Scores in Taiwanese Population and Clinical Cohorts 作者: 柯怡瑄; Yi-Syuan Ke 摘要: 癲癇影響全球逾五千萬人，癲癇因其病因多樣、臨床症狀複雜，導致診斷過程充滿挑戰。儘管全基因體關聯研究 (GWAS) 已識別出多個與常見癲癇相關的基因座，基於這些變異建構的多基因風險分數 (PRS) 逐漸被視為潛在的輔助診斷工具。然而，現有癲癇PRS研究主要集中於歐洲族群，其在非歐洲族群的預測效能仍缺乏驗證。 本研究評估了癲癇 PRS 在三個招募方式與表型定義各異的台灣人群資料 中之預測與分層效能。這些目標樣本分別是：台灣人體生物資料庫 (TWB) 連結全民健康保險研究資料庫 (NHIRD) 的104,680名參與者，其中依ICD診斷碼與抗癲癇藥物紀錄判定915例癲癇病例；台灣精準醫療計畫 (TPMI) 的464,596名參與者，其中依相同標準定義識別5,609例癲癇病例；以及Epi25聯盟之Epi25-TWN樣本的1,140名參與者，內含699例由臨床醫師精細診斷之癲癇個案。接著利用國際抗癲癇組織 (ILAE) 最新的GWAS數據 (族群組成：85% 歐洲、6% 亞洲、9% 非洲)，計算了所有癲癇 (All Epilepsy)、遺傳性全面性癲癇 (GGE) 及局灶性癲癇 (FE) 在這三個樣本中的PRS。隨後，透過終身風險模型、發病年齡分層分析及全表型關聯分析 (PheWAS)，全面評估其預測與風險分層能力。 研究結果顯示GGE PRS在Epi25-TWN樣本中效應最強 (每標準差OR = 1.65, R^2 = 3.2%)，與GGE具高遺傳度 (~40%) 的預期相符，但在兩個大型人口型樣本（TWB-NHIRD與TPMI）中卻僅有相對微弱的效應 (OR = 1.09¬¬–1.13 , R^2 = 0.14–0.22%)。相較之下，FE PRS在三個世代中的效應較¬為一致，但整體屬中 (OR = 1.02–¬1.16, R^2 = 0.01–0.31%）。在風險分層方面，Epi25-TWN中的效果亦最為顯著，GGE PRS排名前5% 高的個體，其患GGE風險為其餘個體的四倍以上，然而相同效應在 TWB-NHIRD 和 TPMI中則較為溫和，風險為其餘個體的不到1.5倍。此外，研究結果支持較高的PRS分數與發病年齡呈負相關，此現象在GGE中尤為顯著，進一步分析發現，將分析限制於較早發病的患者時，GGE PRS在PheWAS中呈現出與癲癇相關的特異訊號。 這項研究是目前針對非歐洲族群規模最大的癲癇 PRS 研究，發現目標樣本研究設計與表型精確度對於PRS預測力的顯著影響，並突顯未來需整合更全面的表型數據與更多元族群的癲癇GWAS，方能顯著提升PRS在跨族群的解釋力與臨床應用潛力。; Epilepsy affects over 50 million people worldwide, and its diagnosis remains challenging due to heterogeneous etiologies and diverse clinical presentations. Genome-wide association studies (GWAS) have identified dozens of loci for common epilepsies, and polygenic risk score (PRS) derived from these findings have been proposed as tools to estimate genetic susceptibility and support diagnosis. However, most epilepsy PRS studies have focused on European populations, and their performance in non-European cohorts is largely unexplored. We assessed the predictive and stratification performance of epilepsy PRSs across three Taiwanese cohorts with distinct recruitment and phenotyping strategies. These cohorts included: 104,680 individuals from the community-based Taiwan Biobank (TWB) linked to 20 years of National Health Insurance Research Database (NHIRD), with 915 epilepsy cases identified using ICD codes and antiseizure medication usage; 464,596 individuals from the multi-hospital-based Taiwan Precision Medicine Initiative (TPMI), including 5,609 cases defined using the same criteria; and 1,140 individuals (700 cases) from the Epi25-TWN cohort (Epi25 consortium) with detailed clinical phenotyping. PRSs for all epilepsy, generalized epilepsy (GGE), and focal epilepsy (FE) were derived from the most recent ILAE GWAS summary statistics (85% European, 6% Asian, 9% African). Predictive performance was evaluated using lifetime risk models, age-of-onset stratification analyses and phenome-wide association study (PheWAS). The GGE PRS showed the strongest effect in Epi25-TWN (OR = 1.65 per SD, R^2 = 3.2%) but weaker effects in the two population-based cohorts, TWB-NHIRD and TPMI (OR = 1.09–1.13, R^2 = 0.14–0.22%). In contrast, the FE PRS yielded more consistent but modest associations across cohorts (OR = 1.02–1.16, R^2 = 0.01–0.31%). Risk stratification was most effective in Epi25-TWN, where individuals in the top 5% of GGE PRS showed over fourfold higher odds of GGE. However, enrichment in the TWB-NHIRD and TPMI was modest, less than 1.5-fold. Furthermore, higher PRS was associated with an earlier age of onset across all three cohorts, especially for GGE, with epilepsy-specific signals for GGE PRS emerging in PheWAS when analyses were restricted to earlier-onset patients. This study represents the largest evaluation of epilepsy PRS performance in non-European populations to date. Our findings highlight how cohort design and phenotyping accuracy significantly influence PRS predictive utility, underscoring the critical need for richer phenotypic data and more diverse epilepsy GWAS to enhance cross-population portability and clinical applicability. 2025-01-01T00:00:00Z http://tdr.lib.ntu.edu.tw/jspui/handle/123456789/99896 標題: 臺灣大腸直腸癌的全外顯子定序研究; A Whole-Exome Sequencing Study of Colorectal Cancer inTaiwan 作者: 施詠齡; Yong-Ling Shih 摘要: 研究背景與動機 根據台灣衛生福利部國民健康署的資料，大腸直腸癌在台灣是最常見的癌症之一現居所有癌症發生率的第二名，死亡率則在癌症相關疾病中排名第三。現今台灣關於大腸直腸癌在基因方面的研究大多是對特定基因組進行分析，但這種研究沒有全外顯子定序方法所分析的結果全面，因此本研究旨在使用台大醫院在大腸直腸癌患者中收集到的全外顯子定序的資料來進行分析，以建構台灣患者的基因突變圖譜及找到特有突變事件，同時也會分析台灣病患的突變特徵。此外，研究中亦會使用全外顯子定序分析過程中產生的檔案來進行拷貝數變異分析以更全面的觀察到台灣大腸直腸癌患者不同染色體區域的拷貝數增益或缺失的情況。 研究方法 研究中所使用的資料來自於台大醫院從209位大腸直腸癌病患中所採集的全外顯子定序資料來進行分析，使用的背景基因體版本為hg19。過程中首先將使用BWA-MEM、Samtools、MarkDuplicates、ApplyBQSR等工具將讀段進行比對並排序，並移除重複讀段及使用常見變異資料庫來進行校正。接著，使用GATK的LearnReadOrientationModel、GetPileupSummaries、CalculateContamination等工具判斷是否有假陽性的情形，並使用Mutect2進行正常樣本與腫瘤樣本的配對變異檢測。最後用FilterMutectCalls與SelectVariants將不合格的突變進行過濾。在突變特徵使用的是SigProfilerExtractor來進行分析。此外，在拷貝數變異的過程中，則是用CollectReadCounts、DenoiseReadCounts及ModelSegments等工具先擷取出區間的讀段數量，再使用正常樣本的參考面板將腫瘤樣本的檔案進行標準化及降噪，最後再找出可能的拷貝數變異片段。 研究結果 在COSMIC v70中關於大腸直腸癌收錄筆數≧100的20個常見突變事件中，研究中所捕捉到的常見大腸直腸癌的突變事件共有15個。包含KRAS（rs112445441_p.G13D、rs12193529_p.G12V/ p.G12D、rs12193530_p.G12V/p.G12S）、TP53（rs28934574_p.R282W、rs28934576_p.R273L、rs121913343_ p.R273C、rs11540652_ p.R248Q、rs28934575_ p.G245S、rs28934578_ p.R175H）、PIK3CA（rs12193213_ p.E542K、rs12193279_ p.H1047R）、APC（rs12193332_ p.R1450X）、BRAF（rs113488022_ p.V600E）。因此在本研究中共捕捉到了75%的常見突變事件。此外，在KRAS、APC、OR2T2、KRT76、ACVR2A、PLEKH6、RNF43、DOCK3上有10個突變事件是在COSMIC v70中關於大腸直腸癌收錄筆數較少，但在研究樣本中突變比例較高的突變事件，可能是台灣大腸直腸癌病患特有的突變事件。 在突變特徵的結果中，主要涵蓋的有年齡、DNA聚合酶ε催化次單元突變、DNA錯配修復缺陷、微衛星不穩定性、DNA聚合酶δ1基因突變、同源重組為基礎的DNA損傷修復功能缺陷、馬兜鈴酸暴露，這些應該是台灣大腸直腸癌病患較常見的致癌原因。 拷貝數變異的結果中，第7、8、13、18、20號染色體有較高的樣本變異比例，其中在第8號染色體前段及第18號染色體為缺失，第7、8後半段、13、20號染色體為增益，且這些區域在僅留下拷貝數≧2.5或拷貝數≦1.5的結果時也同樣明顯。而在僅留下拷貝數≧3或拷貝數≦1的結果時，第13及20號染色體仍有相對較高的樣本增益比例。 討論與結論 研究結果中涵蓋了75%的常見大腸直腸癌突變，發現了10個可能是台灣大腸直腸癌病患特有的突變事件，並根據突變特徵的結果找出了可能的致癌原因。同時，在拷貝數變異分析中找出了台灣大腸直腸癌患者有較明顯拷貝數變異的區域，期望這些結果能在大腸直腸癌患者的治療成效與預後上有所助益。; Background and Motivation According to data from the Health Promotion Administration of Taiwan's Ministry of Health and Welfare, colorectal cancer is one of the most common cancers in Taiwan, currently ranking second in incidence among all cancers and third in cancer-related mortality. Most genetic studies on colorectal cancer in Taiwan to date have focused on specific gene sets; however, such studies lack the comprehensiveness of whole-exome sequencing (WES). Therefore, this study aims to analyze WES data collected from colorectal cancer patients at National Taiwan University Hospital to identify the mutational landscape and unique mutation events specific to Taiwanese patients. In addition, the study will also investigate the mutational signatures present in these patients. Furthermore, copy number variation (CNV) analysis will be conducted using data generated during the WES process to provide a more comprehensive understanding of chromosomal gains and losses in different regions among Taiwanese colorectal cancer patients. Research Methods The data used in this study were obtained from WES of 209 colorectal cancer patients at National Taiwan University Hospital, using the hg19 reference genome. First aligning and sorting the sequencing reads using tools such as BWA-MEM, Samtools, MarkDuplicates, and ApplyBQSR. Duplicated reads were removed, and data were calibrated using common variant databases. Next, using tools such as GATK's LearnReadOrientationModel, GetPileupSummaries, and CalculateContamination were used to identify potential false positives. Variants were then detected by pairing normal and tumor samples using Mutect2. FilterMutectCalls and SelectVariants were applied to filter out variants that did not meet quality standards. For mutation signatures, the analysis was performed using SigProfilerExtractor. In addition, for CNV analysis, tools such as CollectReadCounts, DenoiseReadCounts, and ModelSegments were employed to first extract the read count data from genomic regions. Tumor samples were normalized and denoised using reference panels from normal samples, and potential CNV segments were identified Research Results Among the 20 most common mutations in colorectal cancer recorded in COSMIC v70 with ≥100 entries, 15 common mutations were captured in this study. These include KRAS（rs112445441_p.G13D、rs12193529_p.G12V/ p.G12D、rs12193530_p.G12V/p.G12S）、TP53（rs28934574_p.R282W、rs28934576_p.R273L、rs121913343_ p.R273C、rs11540652_ p.R248Q、rs28934575_ p.G245S、rs28934578_ p.R175H）、PIK3CA（rs12193213_ p.E542K、rs12193279_ p.H1047R）、APC（rs12193332_ p.R1450X）、BRAF（rs113488022_ p.V600E）. Thus, 75% of the common mutations were captured in this study. Additionally, 10 mutations with higher variant frequencies in the study samples were found in KRAS, APC, OR2T2, KRT76, ACVR2A, PLEKH6, RNF43, and DOCK3. These mutations were less frequently recorded in COSMIC v70 for colorectal cancer, suggesting they may be unique to Taiwan colorectal cancer patients . The mutation signature results primarily reflect mutational processes associated with aging, mutations in the catalytic subunit of DNA polymerase epsilon (POLE), DNA mismatch repair deficiency (MMRd), microsatellite instability (MSI), mutations in the DNA polymerase delta 1 (POLD1) gene, defects in homologous recombination-based DNA damage repair (HR repair), and exposure to aristolochic acid. These are likely the major carcinogenic mechanisms of colorectal cancer in Taiwanese patients. Based on the CNV results, chromosomes 7, 8, 13, 18, and 20 showed a higher proportion of samples with variations. Specifically, losses were observed on the proximal region of chromosome 8 and on chromosome 18, while gains were observed on the distal region of chromosome 8, as well as on chromosomes 7, 13, and 20. These regions remained prominent even when applying a stricter threshold of copy number ≥2.5 or ≤1.5. Furthermore, under a more stringent threshold of copy number ≥3 or ≤1, chromosomes 13 and 20 still exhibited relatively high frequencies of copy number gains. Discussion and Conclusion The study covered 75% of common colorectal cancer mutations and identified 10 potentially Taiwan-specific mutations. Based on the mutation signature analysis, possible carcinogenic mechanisms were inferred. Through copy number variation analysis, we identified genomic regions with notable alterations in Taiwanese colorectal cancer patients. These findings may provide insights that contribute to enhancing treatment efficacy and prognosis. 2025-01-01T00:00:00Z