Deep Learning Time-Series Prediction of Long-Term Growth Patterns of Pulmonary Ground-Glass Nodules Using Serial CT

Overview

Pulmonary ground-glass nodules (GGNs) are commonly found on chest CT scans. Some stay stable for years, while others slowly or rapidly turn into lung cancer. Doctors currently follow these nodules with repeated CT scans, but it is difficult to tell ahead of time which nodules will progress, how fast they will progress, and which ones can be safely monitored rather than immediately treated. This observational study aims to develop and validate an artificial intelligence (AI) model that uses each patient's series of CT scans over time to predict the long-term growth behavior of a GGN. The research team will collect three retrospective single-center cohorts from Peking University People's Hospital (a development cohort and two internal test cohorts, one from surgically resected patients and one from non-operated patients followed by serial CT) as well as a prospective multi-center validation cohort enrolled after the AI model is locked. For every patient, each GGN is automatically segmented in three dimensions on every CT scan. A deep learning model extracts imaging features at each timepoint and feeds the sequence of features, together with the actual times between scans, into a time-aware sequence model. The model is trained to predict (i) whether the nodule will show radiological progression at 1, 3, and 5 years after baseline, and (ii) which of four long-term growth patterns the nodule will follow: stable, slow progression, slow-then-rapid progression, or rapid progression. In patients who were ultimately resected, the histopathological diagnosis serves as a secondary reference standard. This is an observational study. No experimental treatment is given. All CT scans and clinical visits are part of routine clinical care.

Pulmonary ground-glass nodules (GGNs), including pure ground-glass nodules (pGGN) and mixed ground-glass nodules (mGGN), span a biological spectrum from atypical adenomatous hyperplasia and adenocarcinoma in situ to invasive lung adenocarcinoma. Current management guidelines (Fleischner Society, BTS, NCCN) rely primarily on cross-sectional CT features (diameter, density, consolidation-to-tumor ratio); these features do not capture the non-linear long-term behavior of GGNs. Long-term cohorts show that a meaningful fraction of GGNs remain indolent for years and then accelerate, demonstrating the limits of single-timepoint assessment. Existing GGN management algorithms emphasize cross-sectional features measured on a single CT scan. Cross-sectional features alone are known to lose discriminative performance at longer (3-5 year) prediction horizons, where the relevant signal is increasingly carried by how the nodule changes over time rather than how it looks at any single moment. The present study is designed around this observation. It shifts the modeling target from static single-CT classification to a long-term spatiotemporal deep-learning framework that explicitly encodes the full trajectory of the nodule across the entire available serial-CT record. This study uses a mixed retrospective-prospective design and is reported under TRIPOD guidance. Four cohorts are pre-specified: * Development Cohort (Cohort 1): n ≈ 2,700. Surgically resected patients at Peking University People's Hospital from January 2007 to June 2025, with ≥ 2 pre-operative thin-slice chest CT scans available for the resected GGN. * Internal Test Cohort A (Cohort 2): n ≈ 350. Surgically resected patients at Peking University People's Hospital from July 2025 to January 2026, with ≥ 2 pre-operative thin-slice chest CT scans available for the resected GGN. * Internal Test Cohort B (Cohort 3): n ≈ 1,200. Non-operated patients at Peking University People's Hospital from January 2020 to December 2025, with ≥ 3 thin-slice chest CT scans of the target GGN. * Prospective External Validation Cohort (Cohort 4): n ≈ 500. Patients (operated or non-operated) prospectively enrolled at Peking University People's Hospital and additional co-sites after the AI model is locked, with serial thin-slice chest CT performed as part of routine surveillance. This cohort serves as independent external (in-time and multi-center) validation and is not used for any model training or hyperparameter tuning. After resampling to 1 × 1 × 1 mm and lung-window normalization, serial CTs are spatiotemporally registered (rigid + deformable) to baseline. For each timepoint, two regions of interest (ROIs) are derived for every target GGN: (i) the intratumoral ROI, defined as the full 3D extent of the nodule itself, automatically segmented using a pre-trained 3D U-Net with expert review; and (ii) the peritumoral ROI, defined as the shell of lung parenchyma extending 5 mm outward from the segmented nodule boundary, with intervening vessels, airways, and pleural surfaces masked out, to capture the perinodular lung microenvironment around the nodule. Radiomic features (per the Image Biomarker Standardization Initiative, IBSI) are extracted separately from each ROI, and deep features are extracted by a 3D convolutional neural network (e.g., 3D ResNet) applied independently to the intratumoral and peritumoral volumes. The two feature streams are concatenated at each timepoint to capture both intra-lesional heterogeneity and the perilesional microenvironment, which prior work from our team has shown to be informative. The ordered sequence of per-timepoint feature vectors, together with the actual inter-scan time intervals, is fed into a time-aware sequence model (LSTM, GRU, or continuous-time Transformer) to model growth dynamics. Clinical covariates (age, sex, smoking history, family history of malignancy, emphysema score) are fused into a multimodal network. Temporal-attention maps and Shapley values are used for interpretability, with separate attribution reported for the intratumoral and peritumoral feature streams. Model performance is reported with AUC and 95% CI for binary 1-, 3-, and 5-year progression endpoints, with DeLong testing against a single-baseline-CT radiomics benchmark; multi-class performance for the four-pattern trajectory task is reported with weighted / macro / micro F1 plus per-class precision and recall. Class imbalance, in particular the slow-then-rapid subgroup, is addressed with ADASYN oversampling and focal loss. For the subset of patients ultimately resected (Cohorts 1, 2, and the operated subset of Cohort 4), histopathological diagnosis serves as a secondary reference standard against which the imaging endpoints and the AI-predicted trajectory are compared.