Explainable machine learning integrating CT-derived intratumoral heterogeneity for histological subtype prediction in pure ground-glass lung adenocarcinoma: multicenter validation and transcriptomic characterization
5 affiliations
- 1Department of Ultrasound, Wuhan Fourth Hospital, Wuhan, China
- 2Hengyang Medical School, University of South China, Hengyang, Hunan, China
- 3Department of Radiology, The Central Hospital of Xiangtan (The Affiliated Hospital of Hunan University), Xiangtan, Hunan, China
- 4Department of Radiology, The Affiliated Hospital of Southwest Medical University, Luzhou, Sichuan, China
- 5Liuyang Traditional Chinese Medicine Hospital, Changsha, Hunan, China
Abstract
Background: Preoperative differentiation of lepidic-predominant adenocarcinoma from other histological subtypes in lung adenocarcinoma presenting as pure ground-glass nodules remains challenging. We evaluated whether a CT-derived intratumoral heterogeneity score provides complementary predictive and biological information beyond routine clinical and CT features.
Methods: This retrospective multicenter study included 533 patients, divided into a training cohort (n = 373) and a held-out validation cohort (n = 160). Twelve supervised algorithms were compared exclusively within the training cohort using stratified five-fold cross-validation and grid-search optimization; predictors, hyperparameters, and the decision threshold were locked before the held-out cohort was evaluated once. Incremental value of the ITH score and leave-one-center-out internal–external validation were additionally assessed. Segmentation reproducibility was evaluated in 100 lesions. An independent radiogenomic cohort of 43 pathological Tis/T1 lung adenocarcinomas was analyzed by differential-expression and Hallmark gene-set enrichment analyses after median ITH-score stratification.
Results: XGBoost achieved the highest training-set cross-validation AUC (0.908 ± 0.021) and was selected before held-out evaluation. In the validation cohort, the AUC was 0.919 (95% CI, 0.870–0.968), with accuracy 0.850, precision 0.857, sensitivity 0.545, specificity 0.966, and Brier score 0.098. Adding the ITH score to clinical and CT features increased AUC from 0.867 to 0.919 (ΔAUC = 0.052; DeLong P = 0.012), increased PR-AUC from 0.731 to 0.824, and reduced Brier score from 0.126 to 0.098. Leave-one-center-out AUCs ranged from 0.824 to 0.882. The ITH score showed good-to-excellent repeat-segmentation reproducibility. In the radiogenomic cohort, a high ITH score was associated with coordinated enrichment of G2M checkpoint (NES = 1.86, FDR <0.001), E2F targets (NES = 1.86, FDR <0.001), and epithelial–mesenchymal transition (NES = 1.55, FDR = 0.001).
Conclusion: CT-derived ITH provided incremental information for histological subtype prediction and was associated with proliferation-, cell-cycle-, and EMT-related transcriptional programs. The moderate sensitivity at the prespecified operating point and the retrospective design warrant prospective external validation before clinical implementation.
Related
- Multi-b-value diffusion-weighted imaging-based habitat imaging for preoperative prediction of histological grade and prognosis in soft tissue sarcoma: focusing on intratumoral heterogeneity
- Machine learning models integrating 2D and 3D intratumoral heterogeneity features from multi-sequence MRI: accurate prediction of perineural invasion in pancreatic ductal adenocarcinoma
- MRI-based intratumoral and peritumoral heterogeneity for predicting intracranial progression in EGFR-mutant lung adenocarcinoma patients with treatment-naive brain metastases receiving first-line third-generation EGFR-TKIs.
- Topologically distinct intratumoral heterogeneity scores for predicting high-risk pathological grades in invasive lung adenocarcinoma: A multicenter study across four institutions
- Senescence-Related Transcriptomic Analysis and Functional Characterization of ANGPTL4 in Lung Adenocarcinoma
BibTeX
@article{deng2026explainable,
title = {{Explainable machine learning integrating CT-derived intratumoral heterogeneity for histological subtype prediction in pure ground-glass lung adenocarcinoma: multicenter validation and transcriptomic characterization}},
author = {Ying Deng and Jiajun Liu and Dinghua Li and Bingyi Zhang and Bin Xiang and Wanyin Qi and Sanhong Zhang},
journal = {Frontiers in Molecular Biosciences},
year = {2026},
month = oct,
doi = {10.3389/fmolb.2026.1958127},
url = {https://doi.org/10.3389/fmolb.2026.1958127},
note = {CC BY 4.0},
}
RIS
TY - JOUR
TI - Explainable machine learning integrating CT-derived intratumoral heterogeneity for histological subtype prediction in pure ground-glass lung adenocarcinoma: multicenter validation and transcriptomic characterization
AU - Ying Deng
AU - Jiajun Liu
AU - Dinghua Li
AU - Bingyi Zhang
AU - Bin Xiang
AU - Wanyin Qi
AU - Sanhong Zhang
PY - 2026
DA - 2026/10/12
JO - Frontiers in Molecular Biosciences
DO - 10.3389/fmolb.2026.1958127
UR - https://doi.org/10.3389/fmolb.2026.1958127
AB - Background: Preoperative differentiation of lepidic-predominant adenocarcinoma from other histological subtypes in lung adenocarcinoma presenting as pure ground-glass nodules remains challenging. We evaluated whether a CT-derived intratumoral heterogeneity score provides complementary predictive and biological information beyond routine clinical and CT features. Methods: This retrospective multicenter study included 533 patients, divided into a training cohort (n = 373) and a held-out validation cohort (n = 160). Twelve supervised algorithms were compared exclusively within the training cohort using stratified five-fold cross-validation and grid-search optimization; predictors, hyperparameters, and the decision threshold were locked before the held-out cohort was evaluated once. Incremental value of the ITH score and leave-one-center-out internal–external validation were additionally assessed. Segmentation reproducibility was evaluated in 100 lesions. An independent radiogenomic cohort of 43 pathological Tis/T1 lung adenocarcinomas was analyzed by differential-expression and Hallmark gene-set enrichment analyses after median ITH-score stratification. Results: XGBoost achieved the highest training-set cross-validation AUC (0.908 ± 0.021) and was selected before held-out evaluation. In the validation cohort, the AUC was 0.919 (95% CI, 0.870–0.968), with accuracy 0.850, precision 0.857, sensitivity 0.545, specificity 0.966, and Brier score 0.098. Adding the ITH score to clinical and CT features increased AUC from 0.867 to 0.919 (ΔAUC = 0.052; DeLong P = 0.012), increased PR-AUC from 0.731 to 0.824, and reduced Brier score from 0.126 to 0.098. Leave-one-center-out AUCs ranged from 0.824 to 0.882. The ITH score showed good-to-excellent repeat-segmentation reproducibility. In the radiogenomic cohort, a high ITH score was associated with coordinated enrichment of G2M checkpoint (NES = 1.86, FDR <0.001), E2F targets (NES = 1.86, FDR <0.001), and epithelial–mesenchymal transition (NES = 1.55, FDR = 0.001). Conclusion: CT-derived ITH provided incremental information for histological subtype prediction and was associated with proliferation-, cell-cycle-, and EMT-related transcriptional programs. The moderate sensitivity at the prespecified operating point and the retrospective design warrant prospective external validation before clinical implementation.
SN - 2296-889X
ER -