Pretreatment CT-based 3D deep learning feature model for predicting treatment response and stratifying survival after immunotherapy rechallenge in hepatocellular carcinoma

Immunotherapy rechallenge is considered for patients with hepatocellular carcinoma (HCC) after failure or discontinuation of prior PD-1 inhibitor–based therapy. However, reliable pretreatment tools for patient selection remain lacking. In this multicenter retrospective study, 200 patients with HCC who underwent PD-1 inhibitor rechallenge between January 2020 and June 2025 were consecutively enrolled from two centers. Among them, 157 patients from the development center were stratified-randomly assigned in a 7:3 ratio to a training cohort ( n = 109) and testing cohort ( n = 48), while 43 patients from another center served as an independent external validation cohort. The primary endpoint was objective response according to modified Response Evaluation Criteria in Solid Tumors. Three pretreatment models were developed and compared: a clinical model, a 3D deep learning (DL) feature model, and a combined clinical–3D DL model. Model performance was assessed using receiver operating characteristic and precision–recall analyses, calibration, and decision curve analysis. The ability of the models to stratify overall survival and progression-free survival was also explored. The development cohort included 157 patients (training, n = 109; testing, n = 48), of whom 26 (16.56%) achieved an objective response. Among the three models, the combined clinical–3D DL LR model showed favorable overall performance, with AUCs of 0.881, 0.853, and 0.804 and Brier scores of 0.088, 0.095, and 0.130 in the training, testing, and external validation cohorts, respectively. In the external validation cohort, the combined model achieved an accuracy of 0.814, sensitivity of 0.556, specificity of 0.882, and positive predictive value of 0.556. It significantly outperformed the clinical model in both the training ( P = 0.007) and testing ( P = 0.028) cohorts and showed favorable calibration and clinical net benefit. The combined clinical–3D DL model also significantly stratified patients according to both PFS (HR = 0.54, P = 0.008) and OS (HR = 0.55, P = 0.031) and achieved higher discriminative performance than conventional clinical staging and scoring systems. SHAP analysis identified AFP as the most influential predictor, together with multiple arterial- and venous-phase 3D DL features. Pretreatment models integrating clinical variables with CECT-derived 3D DL features showed promising performance for noninvasive prediction of response and survival stratification after PD-1 inhibitor rechallenge in HCC. These findings suggest that the integrated clinical–3D DL model may serve as a potential decision-support tool for patient selection and risk stratification in clinical practice. Further prospective multicenter validation is warranted before clinical implementation.

Authors

Institutions

Publication Details

Journal
BMC Medical Informatics and Decision Making
Published
2026-11-14
DOI
https://doi.org/10.1186/s12911-026-03978-0
Primary Topic
Radiomics and Machine Learning in Medical Imaging
Type
article
Field-Weighted Citation Impact
0.00
Controls
|||
ALL TIME
JAN
FEB
MAR
APR
MAY
JUN
JUL
AUG
SEP
OCT
NOV
article

Pretreatment CT-based 3D deep learning feature model for predicting treatment response and stratifying survival after immunotherapy rechallenge in hepatocellular carcinoma

Zenghong Lu, Chaowen Ye, Zheng Li, Lujun Zhang et al.
2 links
BMC Medical Informatics and Decision Making
Radiomics and Machine Learning in Medical Imaging
article

Pretreatment CT-based 3D deep learning feature model for predicting treatment response and stratifying survival after immunotherapy rechallenge in hepatocellular carcinoma

Zenghong Lu, Chaowen Ye, Zheng Li, Lujun Zhang, Rongzhong Huang
article en

Abstract

Immunotherapy rechallenge is considered for patients with hepatocellular carcinoma (HCC) after failure or discontinuation of prior PD-1 inhibitor–based therapy. However, reliable pretreatment tools for patient selection remain lacking. In this multicenter retrospective study, 200 patients with HCC who underwent PD-1 inhibitor rechallenge between January 2020 and June 2025 were consecutively enrolled from two centers. Among them, 157 patients from the development center were stratified-randomly assigned in a 7:3 ratio to a training cohort ( n = 109) and testing cohort ( n = 48), while 43 patients from another center served as an independent external validation cohort. The primary endpoint was objective response according to modified Response Evaluation Criteria in Solid Tumors. Three pretreatment models were developed and compared: a clinical model, a 3D deep learning (DL) feature model, and a combined clinical–3D DL model. Model performance was assessed using receiver operating characteristic and precision–recall analyses, calibration, and decision curve analysis. The ability of the models to stratify overall survival and progression-free survival was also explored. The development cohort included 157 patients (training, n = 109; testing, n = 48), of whom 26 (16.56%) achieved an objective response. Among the three models, the combined clinical–3D DL LR model showed favorable overall performance, with AUCs of 0.881, 0.853, and 0.804 and Brier scores of 0.088, 0.095, and 0.130 in the training, testing, and external validation cohorts, respectively. In the external validation cohort, the combined model achieved an accuracy of 0.814, sensitivity of 0.556, specificity of 0.882, and positive predictive value of 0.556. It significantly outperformed the clinical model in both the training ( P = 0.007) and testing ( P = 0.028) cohorts and showed favorable calibration and clinical net benefit. The combined clinical–3D DL model also significantly stratified patients according to both PFS (HR = 0.54, P = 0.008) and OS (HR = 0.55, P = 0.031) and achieved higher discriminative performance than conventional clinical staging and scoring systems. SHAP analysis identified AFP as the most influential predictor, together with multiple arterial- and venous-phase 3D DL features. Pretreatment models integrating clinical variables with CECT-derived 3D DL features showed promising performance for noninvasive prediction of response and survival stratification after PD-1 inhibitor rechallenge in HCC. These findings suggest that the integrated clinical–3D DL model may serve as a potential decision-support tool for patient selection and risk stratification in clinical practice. Further prospective multicenter validation is warranted before clinical implementation.

BMC Medical Informatics and Decision Making
Second Affiliated Hospital of Chongqing Medical University (CN), Gannan Medical University (CN)
Good health and well-being
Radiomics and Machine Learning in Medical Imaging
AI Navigator

Ask Laika to Summarize, Analyze, and Connect papers live on the map.

Summarize Papers & Methodologies

Extract key findings, datasets, and comparative methods across publications.

Benchmark Rankings & Visual Analytics

Rank top research institutions, authors, funders, topics, and journals by Field-Weighted Citation Impact (FWCI) and paper volume with instant charts.

Connect Distant Disciplines

Bridge topological clusters on the map to find hidden collaborative intersections.

Laika access coming soon