PO.BCS01.09 · 生物信息与计算
AI衍生的细胞核形态测量学作为肺腺癌新型预后指标
AI-derived nuclear morphometrics as a novel prognostic indicator in lung adenocarcinoma
作者与单位 Authors & Affiliations
摘要 Abstract
中文摘要
背景：在肺腺癌（LUAD）中，预后判断很大程度上依赖于结构特征，包括主要浸润模式和肿瘤浸润大小。然而，细胞核形态在LUAD中的预后相关性相对而言仍未得到充分探索。本研究使用基于AI的细胞形态分析器，旨在系统量化LUAD的各种细胞核形态测量指标并评估其预后意义。
方法：检索手术切除的LUAD病例（n = 160）的全切片图像，应用CellViT提取核水平的形态测量特征。从生成的轮廓中计算了114项特征——包括面积、边界框面积、凸包面积、Feret直径、最大长/短轴长度、纵横比、圆度、形状因子、离心率、紧凑度、坚实度、方位角和分形维数。对于方位相关特征，使用方位方差和方位方差熵生成病例水平特征；对于所有其他变量，使用中位数、IQR（四分位距）、基于IQR的变异系数、直方图熵和分位数（Q10、Q20、Q80和Q90）生成病例水平特征。使用这些病例水平特征拟合疾病特异性生存（DSS）和无复发生存（RFS）的单变量Cox模型。
结果：多项细胞核轮廓特征，包括最大Feret直径、周长、长轴长度、边界框面积和凸包面积，是DSS（107/114，93.9%）和RFS（28/114，24.6%）的显著预后因素。
讨论：这些发现表明细胞核形态测量学在LUAD中可能具有预后意义。将基于AI的核特征纳入现有分级或风险分层框架可能提高预后精度并减少对主观视觉评估的依赖。AI驱动的核形状分析有望成为肺癌病理学中的互补性生物标志物。
LUAD各种细胞核形态测量指标的单变量分析 疾病特异性生存特征 校正P值 无复发生存特征 校正P值 最大Feret直径_q90 <0.001 长轴长度_q90 <0.001 周长_q90 <0.001 最大Feret直径_q90 <0.001 长轴长度_q90 <0.001 周长_q90 <0.001 面积_bbox_q90 <0.001 面积_bbox_q90 0.001 凸包面积_q90 <0.001 凸包面积_q90 0.003 面积_q90 <0.001 长轴长度_q80 0.004 最大Feret直径_q80 <0.001 最大Feret直径_q80 0.004 长轴长度_q80 <0.001 面积_q90 0.004 分形维数_q20 <0.001 FD_中位数 0.010 分形维数_q10 <0.001 周长_q80 0.011
查看英文原文 English abstract
Background: In lung adenocarcinoma (LUAD), prognostication has largely relied on architectural features, including predominant invasive pattern and tumor invasive size. However, the prognostic relevance of nuclear morphology relatively remains underexplored in LUAD. Using AI-based cell morphology analyzer, this study aimed to systematically quantify various nuclear morphometrics of LUAD and evaluate its prognostic significance.
Method: Whole slide images of surgically resected LUAD cases (n = 160) were retrieved, and CellViT was applied to extract nuclear-level morphometric features. From the generated contours, 114 features-including area, bounding-box area, convex-hull area, Feret diameter, maximum major/minor axis length, aspect ratio, circularity, form factor, eccentricity, compactness, solidity, orientation angle, and fractal dimension-were computed. Case-level features were generated for orientation-related features using orientation variance and entropy of orientation variance, and for all other variables using the median, IQR (interquartile range), IQR-based coefficient of variation, histogram entropy and quantiles (Q10, Q20, Q80, and Q90). Univariate Cox models for disease-specific survival (DSS) and recurrence-free survival (RFS) were fitted using these case-level features.
Results: Multiple nuclear contour features, including maximum Feret diameter, perimeter, major axis length, bounding-box area, and convex-hull area, were significant prognostic factors for both DSS (107/114, 93.9%) and RFS (28/114, 24.6%).
Discussion: These findings demonstrate that nuclear morphometrics can have prognostic implication in LUAD. Incorporating AI-based nuclear features into current grading or risk-stratification frameworks may improve prognostic precision and reduce dependence on subjective visual assessment. AI-driven nuclear shape profiling holds promise as a complementary biomarker in lung cancer pathology.
Univariable analysis of various nuclear morphometrics of LUAD Disease specific survival feature Adjusted P-value Recurrence free survival feature Adjusted P-value Maximum Feret diameter_q90 <0.001 Major axis length_q90 <0.001 Perimeter_q90 <0.001 Maximum Feret diameter_q90 <0.001 Major axis length_q90 <0.001 Perimeter_q90 <0.001 Area_bbox_q90 <0.001 Area_bbox_q90 0.001 Area_convex_q90 <0.001 Area_convex_q90 0.003 Area_q90 <0.001 Major axis length_q80 0.004 Maximum Feret diameter_q80 <0.001 Maximum Feret diameter_q80 0.004 Major axis length_q80 <0.001 Area_q90 0.004 Fractal dimension_q20 <0.001 FD_median 0.010 Fractal dimension_q10 <0.001 Perimeter_q80 0.011
利益披露 Disclosure
B. Ahn, None..
H. Hwang, None..
H. Sung, None..
S. Jang, None..
P. Kim, None..
H. Go, None.