Scholarlink

Purpose <p>Preoperative differentiation of hepatocellular carcinoma (HCC), intrahepatic cholangiocarcinoma (ICC), and combined hepatocellular-cholangiocarcinoma (cHCC-CCA) remains clinically challenging due to overlapping imaging features. This study evaluated whether radiomic features from multiphasic contrast-enhanced CT can simultaneously differentiate these three primary liver cancer subtypes using a bias-aware machine learning framework.</p> Methods <p>A publicly available four-phase contrast-enhanced CT (CECT) dataset of 278 pathologically confirmed patients (94 HCC, 99 ICC, 85 cHCC-CCA) was analyzed. Intratumoral, peritumoral, boundary, and background-liver radiomic features (<i>n</i> = 373) were compared across classes using the Kruskal-Wallis test with false discovery rate correction. An XGBoost classifier was trained within a nested cross-validation framework comparing four feature-selection strategies; the SHAP-based top-K strategy was retained as the primary model. Model significance was confirmed by permutation testing, and misclassified cases were characterized using key discriminative features.</p> Results <p>Of 373 tested features, 316 (84.7%) showed significant intergroup differences after correction, led by arterial-phase liver-parenchyma reference attenuation (<InlineEquation ID="IEq1"> <EquationSource Format="TEX">\(\:{\eta\:}^{2}\)</EquationSource> </InlineEquation>= 0.438). The compact SHAP-based model achieved a nested cross-validated AUC of 0.953 ± 0.030 (out-of-fold macro AUC 0.948; permutation <i>p</i> &lt; 0.001). In a separate ablation analysis, background-liver features alone approached full-model performance, exceeding intratumoral features alone. Of 278 patients, 48 (17.3%) were misclassified, predominantly at the ICC-cHCC-CCA boundary; errors reflected attenuation values shifted toward the confounding class, except in cHCC-CCA, where no examined variable distinguished correctly from misclassified cases.</p> Conclusion <p>These findings are hypothesis-generating and require external validation; radiomic features describing the tumor-liver interface and peritumoral microenvironment showed high cross-validated discriminative performance and, pending prospective validation, may inform non-invasive assessment in cases of diagnostic uncertainty.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Three-class radiomic differentiation of hepatocellular carcinoma, intrahepatic cholangiocarcinoma, and combined hepatocellular-cholangiocarcinoma on multiphasic contrast-enhanced CT

  • 期刊论文

  • 发表日期：

    2026年7月30日

查看全文
  • Yiğit Hasan Arı

摘要

Purpose

Preoperative differentiation of hepatocellular carcinoma (HCC), intrahepatic cholangiocarcinoma (ICC), and combined hepatocellular-cholangiocarcinoma (cHCC-CCA) remains clinically challenging due to overlapping imaging features. This study evaluated whether radiomic features from multiphasic contrast-enhanced CT can simultaneously differentiate these three primary liver cancer subtypes using a bias-aware machine learning framework.

Methods

A publicly available four-phase contrast-enhanced CT (CECT) dataset of 278 pathologically confirmed patients (94 HCC, 99 ICC, 85 cHCC-CCA) was analyzed. Intratumoral, peritumoral, boundary, and background-liver radiomic features (n = 373) were compared across classes using the Kruskal-Wallis test with false discovery rate correction. An XGBoost classifier was trained within a nested cross-validation framework comparing four feature-selection strategies; the SHAP-based top-K strategy was retained as the primary model. Model significance was confirmed by permutation testing, and misclassified cases were characterized using key discriminative features.

Results

Of 373 tested features, 316 (84.7%) showed significant intergroup differences after correction, led by arterial-phase liver-parenchyma reference attenuation ( \(\:{\eta\:}^{2}\) = 0.438). The compact SHAP-based model achieved a nested cross-validated AUC of 0.953 ± 0.030 (out-of-fold macro AUC 0.948; permutation p < 0.001). In a separate ablation analysis, background-liver features alone approached full-model performance, exceeding intratumoral features alone. Of 278 patients, 48 (17.3%) were misclassified, predominantly at the ICC-cHCC-CCA boundary; errors reflected attenuation values shifted toward the confounding class, except in cHCC-CCA, where no examined variable distinguished correctly from misclassified cases.

Conclusion

These findings are hypothesis-generating and require external validation; radiomic features describing the tumor-liver interface and peritumoral microenvironment showed high cross-validated discriminative performance and, pending prospective validation, may inform non-invasive assessment in cases of diagnostic uncertainty.