基於L3影像特徵與機器學習之大腸直腸癌肝轉移患者生存預測：隨機生存森林模型建構

Abstract

大腸直腸癌肝轉移（Colorectal Liver Metastases, CRLM）患者預後異質性高，現行 Fong Clinical Risk Score 僅整合腫瘤負荷指標，未涵蓋骨骼肌、脂肪與骨骼健康度等宿主因素；傳統第三腰椎（Third Lumbar Vertebra, L3）體組成分析多依賴人工標註，難以於臨床落實。本研究自 The Cancer Imaging Archive（TCIA）Colorectal-Liver-Metastases 納入完整案例 168 名患者（排除 CRS 缺失者）作為訓練與內部驗證集，並另取臺大醫院（National Taiwan University Hospital, NTUH）22 名患者為外部驗證集；L3 切面以 TotalSegmentator 自動定位，肌肉與脂肪之多組織分割由既有深度學習身體組成模型產出，並另整合 TotalSegmentator 之 L3 椎體遮罩以量化椎體骨骼（面積與骨骼平均衰減值）及串接以骨骼肌指數為基礎之肌少症自動辨識，生存預測以隨機生存森林（Random Survival Forest, RSF）為主並與 XGBoost Survival Embeddings（XGBSE）於五折交叉驗證下對照，以 SHAP（SHapley Additive exPlanations）提供可解釋性。自動化流程於數十秒內完成單例處理，RSF 於 TCIA 訓練／內部驗證集 C-Index 為 0.9073 ／ 0.7333、外部驗證集 0.7261、五折平均 0.7163 ± 0.013，略優於 XGBSE（0.7013 ± 0.014），內部驗證集 Kaplan-Meier 達 log-rank p = 0.0262，SHAP 顯示臨床風險評分（CRS）、年齡、皮下脂肪面積與新增之椎體骨骼面積為主要貢獻特徵，骨骼平均衰減值等亦進入前段，且內、外部驗證集之主導特徵高度一致，外部驗證集（n=22）則初步顯示與內部驗證一致之區分趨勢。本研究主要成果為建立一套可重現之自動化 L3 體組成分析流程，並以 CRLM 為示範場景初步驗證其方法學可行性；影像端已封裝為桌面端軟體，新增之椎體骨骼量化填補既有工具缺口，初步生存預測結果為「L3 體組成結合機器學習」於癌症預後評估提供方法學可行性證據。受限於樣本規模，本研究之生存預測結果尚不足以支持直接之臨床部署宣稱。

Prognostic heterogeneity in Colorectal Liver Metastases (CRLM) is poorly captured by tumor-burden tools such as the Fong Clinical Risk Score, which omit host factors such as muscle, adipose and bone health, while manual analysis at the third lumbar vertebra (L3) for body composition is difficult to deploy clinically. In this study, 168 complete-case CRLM patients (after excluding cases with missing CRS) from The Cancer Imaging Archive (TCIA) Colorectal-Liver-Metastases dataset were used for training and internal validation and 22 patients from the National Taiwan University Hospital (NTUH) served as the external cohort; L3 was localized with TotalSegmentator, multi-tissue segmentation of muscle and adipose tissue was performed by an existing deep-learning model, and vertebral-bone metrics (bone area and mean CT attenuation) were obtained by quantifying the TotalSegmentator L3 vertebra mask, with an automated sarcopenia classifier appended, while survival prediction used Random Survival Forest (RSF) benchmarked against XGBoost Survival Embeddings (XGBSE) under five-fold cross-validation, with SHapley Additive exPlanations (SHAP) for interpretability. The pipeline processed each case within tens of seconds, RSF reached C-Index = 0.9073 (training), 0.7333 (internal test) and 0.7261 (external test) with a five-fold mean of 0.7163 ± 0.013 slightly outperforming XGBSE (0.7013 ± 0.014), internal Kaplan-Meier stratification reached log-rank p = 0.0262, SHAP identified the Clinical Risk Score, age, subcutaneous adipose area and the newly added vertebral bone area as leading contributors (with mean bone attenuation also in the upper tier) and the dominant features were highly consistent between the internal and external cohorts, and the small external cohort (n=22) showed a preliminary discriminative trend consistent with internal validation. The main contribution is a reproducible automated L3 body-composition workflow demonstrated here on CRLM as the showcase scenario; the imaging components are packaged as a desktop application, the added vertebral-bone quantification fills a gap in existing tools, and the exploratory survival-prediction results provide methodological feasibility evidence for combining L3 body composition with machine learning in cancer prognostication.