2026-10-302026-10-30https://scholars.csmu.edu.tw/handle/123456789/36943Lung cancer the leading cause of cancer-related death worldwide as well as in Taiwan. Moreover, the incidence of lung cancer is increasing, and most of the lung cancer patients were diagnsosed with late-stage disease. Consequently, lung cancer causes an economic burden on the healthcare system. In 2023, it incurred over 22.4 billion in healthcare expenditures. As a result, low-dose computed tomography (LDCT) has been recommended to detect lung cancer at early-stage worldwide. In Western contries, those who smoked composed more than 80% of patients with lung cancer. On the contrary, more than 53% of lung cancer patients are never-smokers in Taiwan. This finding emphasizes the differences in the aetiology between different populations, and the risk factors of lung cancer should be further explored here in Asia. A previous prospective study (TALENT) has proven the importance of family history in the tumorigenesis of lung cancer. Nonetheless, since cancer is often multifactorial, other factors should be explored. Hence, we propose incorporating other variables, such as air pollution, biomarkers, and radiomics, into developing a lung cancer prediction model. We aim to develop a risk prediction model that caters to never-smokers in Taiwan and hope to extrapolate its use to Southeast Asia and Western countries. To build, verify, and refine our risk prediction model, we initiated a multicenter prospective study with four programs. Program 1: Validating non-smoker lung cancer prediction model among Taiwanese population: Integrating environmental and occupational factors. The main aim of this progam is to validate and optimise existing prediction models. We employ machine learning algorithms and incorporate variables including epidemiologic, environmental and occupational factors. We hypothesise that adding GIS-based environmental exposure measurements, and occupational exposure using the job-exposure matrix as a proxy can increase the predictive power of the lung cancer risk model. Program 2: Validation of autoantibody- and genetic prediction model for non-smoker lung cancer. In the past three years, we completed user training for the Early CDT Lung Detection Kit, and “DNA Mass Spectrometer-based Mitochondrial Gene Mutations assay” and confirmed assay quality of autoantibody detection; mitochondrial mutation detection; and TWB 2.0 Array with a small cohort. Eventually, we performed near 2,000 tests for mitochondrial mutation detection and TWB 2.0 array and 1,000 tests for the autoantibody detection assay. In this year, we will finish all of assays for enrolled subjects and calculate the association between biomarkers and lung cancer risk. Program 3: Detection, classification, and prediction of lung cancer risk with CT using deep learning and radiomics. We aim to build an integrated platform for detecting and following up lung nodules that are at a higher risk of developing into lung cancer. To do so, we target solid or part-solid nodules ≥ 6-8mm or GGNs that ≥10mm diameters. Other variables such as volume measurement, morphology, texture, and other CT-radiomics of these nodules should be obtained and incorporated into the risk prediction model along with the modified Lung-RADS system. In addition, we also incoperated AI-CXR in this program. Once built, this prediction model aims to precisely pinpoint participants with high lung cancer risks. Program 4: Optimization and validation of lung cancer risk and probability prediction model: prospective multicenter clinical study. We opted for a case-control study design; we aim to first optimise the risk prediction models from programs 1,2, and 3 using our retrospective cohort. Secondly, we aim to verify the optimised risk prediction models with our prospective cohort. We performed risk stratification for the participants, while biomarkers measurements and LDCT would be conducted for high-risk subjects. The optimized models including biomarkers and image features will be performed to predict the probability of lung cancer occurrence and compare it with the conventional diagnosis methods in low-risk participants. Our final goal is to propose a lung cancer screening strategy that caters to the Taiwanese population.肺癌近年來已成為最主要癌症死亡原因之首位，其發病率也呈現逐年上升的趨勢。112年醫療支出高達224.9億元，為耗費健保費用最多的癌症，其中主要原因是半數患者在診斷時已為晚期。因此，低劑量電腦斷層(LDCT)檢查已然成為全世界最普遍用於肺癌篩檢的工具。然而，西方族群的肺癌有八成以上是因抽菸造成，故其篩查對象主要以抽菸者為主，而在臺灣有53%肺癌是不抽菸者，至今成因仍不明。根據臺灣前瞻性TALENT LDCT研究案結果顯示，肺癌家族史與肺癌發生高度相關，但受限於僅針對少數風險因子收案，未能對空汙、生物標記或影像學特徵納入考量，恐無法精確找出肺癌高風險族群。因此有效整合一個有用的前瞻性驗證與優化多面向不抽菸肺癌風險預測模型對臺灣，甚至東亞或西方不抽菸肺癌都有幫助。藉此，我們提出同時對流行病學資訊、空汙、生物標記及影像學特徵等肺癌預測模式進行一前瞻性確效與優化研究案。為達此目的，將執行4個子計畫對臺灣多面向不抽菸肺癌風險預測模型進行驗證。 子計畫1：統合環境及職業因素之臺灣民眾非抽菸肺癌預測模式驗證。目標納入更全面而精確的流行病學因素，利用機器學習演算法驗證和優化現有預測模型，我們假設肺癌風險模型在國內族群可被驗證，且其預測能力將增加在納入更多人口學變項。在環境因子部分，驗證當前基於空汙、重金屬的不抽菸者肺癌風險預測模型，並以更高解析度之暴露評估因子，優化現有模型中環境和職業參數。 子計畫2：自體抗體與基因生物標記不抽菸肺癌風險預測模型驗證。在計畫執行前三年，我們完成早期肺癌自體抗體檢測的人員訓練與DNA質譜儀粒線體突變檢測的確效，自體抗體檢測、粒線體突變檢測與TWB 2.0晶片的小群組效能評估。最後，完成近2,000例粒線體突變檢測與TWB2.0晶片及1,000例自體抗體篩查檢測。在今年，我們將完成所有收案者的自體抗體與基因生物標記檢測，同時計算生物標記與肺癌風險的關聯性分析。 子計畫3：使用深度學習與影像組學於CT偵測、分類與預測肺癌風險。持續優化LDCT肺結節輔助偵測追蹤平台，並架構以影像醫學為基礎之肺結節相似度評估方法。對於偵測之結節在大於等於6-8mm直徑以上實心或部分實心結節與大於10mm毛玻璃結節將配合Lung-RADS評估並進行CT影像體學(CT radiomic)分析，以尋找較高肺癌風險之受試者。建立AI CXR判讀方式。 子計畫4：肺癌風險與機率預測模型優化與驗證：前瞻性多中心臨床研究。初步建置肺癌風險預測模型雛型，並優化子計畫2、3的生物標記及影像特徵模型，再以前瞻性多中心研究設計進行驗證。首先評估受試者風險等級，篩選出的高風險個案將量測生物標記並接受低劑量胸部電腦斷層檢查，再利用預測個案為肺癌的機率與現行臨床診斷方法及低風險個案進行比較。最後使用本計畫整合的肺癌風險及機率模型的驗證，提出適合臺灣族群的肺癌篩檢流程與策略。肺癌預測模式自體抗體粒線體突變機器學習低劑量電腦斷層影像影像體學肺癌風險篩檢策略lung cancer prediction modelautoantibodiesmictochondria mutationsmachine learningLow dose CT (LDCT)radiomicslung cancer riskscreening strategyProject