Development and validation of a deep learning model integrating ¹⁸F-FDG PET/CT radiomics and clinical features to predict major pathological response in non-small cell lung cancer patients receiving chemoimmunotherapy.
7 affiliations
- 1Guangdong Lung Cancer Institute, Guangdong Provincial People's Hospital (Guangdong Academy of Medical Sciences), Southern Medical University, Guangzhou, China.
- 2School of Artificial Intelligence, Beihang University, Beijing, China.
- 3Department of Radiology, The First Affiliated Hospital of Anhui Medical University, Hefei, China.
- 4Department of Thoracic Surgery, The Third People's Hospital of Chengdu, Chengdu, China.
- 5School of Medicine, South China University of Technology, Guangzhou, China.
- 6PET Center, Department of Nuclear Medicine, Guangdong Provincial People's Hospital (Guangdong Academy of Medical Sciences), Southern Medical University, Guangzhou, China.
- 7Department of Pathology, Guangdong Provincial People's Hospital (Guangdong Academy of Medical Sciences), Southern Medical University, Guangzhou, China.
Abstract
Pathological complete response (pCR) and major pathological response (MPR) are early measures of treatment efficacy with prognostic relevance in patients with non-small cell lung cancer (NSCLC) receiving neoadjuvant chemoimmunotherapy. Individual biomarkers such as programmed cell death-ligand 1 (PD-L1) and tumor mutational burden (TMB), and conventional imaging interpretation, incompletely characterize treatment sensitivity. Integrating clinical biomarkers with radiomics may improve prediction by combining information on host factors and tumor biology with quantitative imaging measures of tumor heterogeneity. We aimed to develop and validate a model predicting pathological response. We retrospectively collected fluorine-18 fluorodeoxyglucose positron emission tomography/computed tomography (18F-FDG PET/CT) images and clinical data from 320 patients with NSCLC at Guangdong Provincial People's Hospital who received at least two cycles of chemoimmunotherapy followed by surgical resection and had postoperative pathological response assessments. Patients were randomly assigned to training (n=256) and validation (n=64) cohorts at an 8:2 ratio. A total of 1,834 CT and 179 PET radiomics features were extracted and integrated with demographic and tumor metabolic parameters, PD-L1 expression, TMB, and peripheral blood biomarkers. Following standardization and within-modality Pearson correlation filtering, features were selected using recursive feature elimination (RFE). A modality-aware attention model was developed to generate patient-specific scores for predicting MPR, including pCR. The model's performance was further validated in an external cohort of 31 patients from two independent hospitals. A total of 351 patients were included: 256 in the training cohort (160 responders and 96 non-responders), 64 in the internal testing cohort (35 responders and 29 non-responders), and 31 in the preliminary external test cohort. Our study proposes a modality-aware attention model, which was constructed upon RFE feature engineering technique. Our model integrates diverse data sources, including nine clinical characteristics, nine CT features, and nine PET features. Our model achieved an area under the receiver operating characteristic curve (AUC) of 0.7467 [95% confidence interval (CI): 0.6826 to 0.8159] in the training set. In the external independent validation cohort, the AUC was 0.7017 (95% CI: 0.5126 to 0.8715), with a sensitivity of 82% and a specificity of 86%. Compared to single-modality models based on clinical features (AUC =0.6931; 95% CI: 0.6232 to 0.7604), the modality-aware attention model significantly improved accuracy. The multimodal model showed potential for predicting MPR, including pCR, in patients with NSCLC receiving chemoimmunotherapy. Larger prospective multicenter studies are needed before clinical implementation.
Related
- MRI-Based Radiomics Model for Pretreatment Prediction of Pathological Lymph Node Metastasis in Neuroendocrine Carcinoma of the Cervix: Dual-Center Development and External Validation.
- Development and internal validation of a multimodal deep learning model integrating imaging, clinical and pathological features to predict false-negatives in non-specific benign lung biopsies: a dual-center study.
- A multimodal deep learning radiomics nomogram integrating clinicopathological features for predicting pathological complete response after neoadjuvant therapy in triple-negative breast cancer: a multicenter retrospective study
- Development and internal validation of a clinicopathological and CT radiomics based prediction model for extracranial distant metastasis after radical resection in early-stage non-small cell lung cancer
- Development of a nomogram integrating contrast-enhanced magnetic resonance imaging radiomics and clinical parameters for differentiating brain metastases of bronchial carcinoma from cerebral tuberculomas.
BibTeX
@article{li2026development,
title = {{Development and validation of a deep learning model integrating ¹⁸F-FDG PET/CT radiomics and clinical features to predict major pathological response in non-small cell lung cancer patients receiving chemoimmunotherapy.}},
author = {Ao Li and Zi-Hao Mao and Hai-Long Wang and Yuan Wang and Qi-Hui Fu and Wen-Chao Feng and Yang Chen and Yin-Ting Hu and Zhi-Hua Liu and Wen-Zhao Zhong and Lei Jiang and Ben-Yuan Jiang},
journal = {Translational lung cancer research},
year = {2026},
month = oct,
doi = {10.21037/tlcr-2026-0739},
url = {https://europepmc.org/article/MED/42999401},
note = {CC BY-NC-ND},
}
RIS
TY - JOUR
TI - Development and validation of a deep learning model integrating ¹⁸F-FDG PET/CT radiomics and clinical features to predict major pathological response in non-small cell lung cancer patients receiving chemoimmunotherapy.
AU - Ao Li
AU - Zi-Hao Mao
AU - Hai-Long Wang
AU - Yuan Wang
AU - Qi-Hui Fu
AU - Wen-Chao Feng
AU - Yang Chen
AU - Yin-Ting Hu
AU - Zhi-Hua Liu
AU - Wen-Zhao Zhong
AU - Lei Jiang
AU - Ben-Yuan Jiang
PY - 2026
DA - 2026/10/23
JO - Translational lung cancer research
DO - 10.21037/tlcr-2026-0739
UR - https://europepmc.org/article/MED/42999401
AB - Pathological complete response (pCR) and major pathological response (MPR) are early measures of treatment efficacy with prognostic relevance in patients with non-small cell lung cancer (NSCLC) receiving neoadjuvant chemoimmunotherapy. Individual biomarkers such as programmed cell death-ligand 1 (PD-L1) and tumor mutational burden (TMB), and conventional imaging interpretation, incompletely characterize treatment sensitivity. Integrating clinical biomarkers with radiomics may improve prediction by combining information on host factors and tumor biology with quantitative imaging measures of tumor heterogeneity. We aimed to develop and validate a model predicting pathological response. We retrospectively collected fluorine-18 fluorodeoxyglucose positron emission tomography/computed tomography (18F-FDG PET/CT) images and clinical data from 320 patients with NSCLC at Guangdong Provincial People's Hospital who received at least two cycles of chemoimmunotherapy followed by surgical resection and had postoperative pathological response assessments. Patients were randomly assigned to training (n=256) and validation (n=64) cohorts at an 8:2 ratio. A total of 1,834 CT and 179 PET radiomics features were extracted and integrated with demographic and tumor metabolic parameters, PD-L1 expression, TMB, and peripheral blood biomarkers. Following standardization and within-modality Pearson correlation filtering, features were selected using recursive feature elimination (RFE). A modality-aware attention model was developed to generate patient-specific scores for predicting MPR, including pCR. The model's performance was further validated in an external cohort of 31 patients from two independent hospitals. A total of 351 patients were included: 256 in the training cohort (160 responders and 96 non-responders), 64 in the internal testing cohort (35 responders and 29 non-responders), and 31 in the preliminary external test cohort. Our study proposes a modality-aware attention model, which was constructed upon RFE feature engineering technique. Our model integrates diverse data sources, including nine clinical characteristics, nine CT features, and nine PET features. Our model achieved an area under the receiver operating characteristic curve (AUC) of 0.7467 [95% confidence interval (CI): 0.6826 to 0.8159] in the training set. In the external independent validation cohort, the AUC was 0.7017 (95% CI: 0.5126 to 0.8715), with a sensitivity of 82% and a specificity of 86%. Compared to single-modality models based on clinical features (AUC =0.6931; 95% CI: 0.6232 to 0.7604), the modality-aware attention model significantly improved accuracy. The multimodal model showed potential for predicting MPR, including pCR, in patients with NSCLC receiving chemoimmunotherapy. Larger prospective multicenter studies are needed before clinical implementation.
SN - 2218-6751
SN - 2226-4477
ER -