Retrospective Multicenter Study of Patient-level T1CE/FLAIR MRI Deep Learning to Predict EGFR/ALK Driver Status in NSCLC Brain Metastases With External Validation and Survival Analysis

Brain Metastases, Non-Small Cell Lung Cancer

Magnetic Resonance Imaging, Deep Learning, Multimodal MRI, T1-weighted Contrast-Enhanced (T1CE), FLAIR, Transformer, EGFR, ALK

This retrospective multicenter observational study aims to develop and externally validate a noninvasive deep learning model based on routine brain MRI to identify actionable driver alterations in patients with non-small cell lung cancer (NSCLC) brain metastases. The model uses contrast-enhanced T1-weighted imaging (T1CE) and FLAIR sequences to classify patients as driver-positive (EGFR mutation and/or ALK rearrangement/fusion) versus driver-negative (EGFR-negative and ALK-negative), using brain metastasis tissue next-generation sequencing as the reference standard. The development and internal validation cohorts are from the National Cancer Center (China). Two independent external test cohorts are used: one from the First Affiliated Hospital of Anhui Medical University (China) and one from a public de-identified dataset hosted by The Cancer Imaging Archive (TCIA). The primary endpoint is the patient-level area under the receiver operating characteristic curve (AUC) in the external test cohorts. Secondary analyses include model calibration and decision-curve analysis to estimate clinical utility, comparisons of 2D/2.5D/3D modeling strategies and multimodal fusion approaches, and exploratory associations between model outputs and overall survival (OS) and progression-free survival (PFS), calculated from the date of brain metastasis surgery to the event or last follow-up (data cutoff: May 1, 2026).

Rationale and Objectives Actionable driver alterations such as EGFR mutations and ALK rearrangements/fusions are key determinants of treatment selection in NSCLC. In patients with brain metastases, tissue acquisition may be limited by surgical risk, lesion location, and time constraints. Routine brain MRI provides rich phenotypic information that may capture imaging correlates of molecular drivers. This study is designed to develop and externally validate a patient-level deep learning model that leverages multimodal MRI (T1CE and FLAIR) to noninvasively identify driver-positive status (EGFR mutation and/or ALK rearrangement/fusion) versus driver-negative status (EGFR-negative and ALK-negative). Study Design and Data Sources This is a retrospective multicenter observational cohort study. Model development (training and internal validation) will be performed using data from the National Cancer Center (China). External validation will be conducted in two independent cohorts: (1) a clinical cohort from the First Affiliated Hospital of Anhui Medical University (China) and (2) a public de-identified cohort obtained from The Cancer Imaging Archive (TCIA). The TCIA cohort is used as an independent test set and is not involved in model training, hyperparameter tuning, or threshold selection. Reference Standard and Driver Definition Driver status will be determined by next-generation sequencing performed on resected brain metastasis tissue. Driver-positive is defined as EGFR mutation and/or ALK rearrangement/fusion detected on brain metastasis tissue testing. Driver-negative is defined as both EGFR-negative and ALK-negative. Imaging Inputs and Preprocessing Eligible patients must have preoperative brain MRI including at minimum T1CE and FLAIR sequences with acceptable image quality. Imaging data will be de-identified and standardized for analysis. Preprocessing will include harmonized spatial resampling to a common voxel spacing, intensity normalization, and co-registration between modalities when needed. Lesion localization/segmentation will be performed using manual, semi-automated, or automated approaches with quality control by trained reviewers, depending on data availability. For patients with multiple brain metastases, lesion-level representations will be aggregated to produce a patient-level prediction using a predefined pooling strategy (e.g., attention pooling or multiple-instance learning). Model Development and External Validation The primary model will use multimodal inputs (T1CE + FLAIR) and a fusion strategy (including transformer-based fusion as a prespecified approach). Comparative analyses will evaluate 2D, 2.5D, and 3D modeling strategies and alternative fusion schemes (e.g., early vs late fusion) under a consistent evaluation framework. All model selection and threshold determination will be completed using the National Cancer Center development data. The finalized model and prespecified thresholds will then be locked and evaluated once in each external cohort without any additional training or recalibration. Outcomes and Statistical Analysis The primary endpoint is discrimination performance assessed by patient-level AUC in the external test cohorts, with 95% confidence intervals. Secondary endpoints include sensitivity, specificity, positive predictive value (PPV), negative predictive value (NPV), calibration (calibration curves and Brier score), and decision-curve analysis to estimate net benefit across clinically relevant thresholds. Subgroup analyses may be performed by center, imaging acquisition characteristics, and single versus multiple metastases. In a subset with follow-up data, exploratory analyses will evaluate associations between model outputs and OS/PFS using Kaplan-Meier methods and Cox proportional hazards models. OS and PFS will be calculated from the date of brain metastasis surgery to death/progression or last follow-up; the data cutoff date is May 1, 2026. Ethics and Privacy This study uses retrospective clinical data that will be de-identified prior to analysis. Institutional review board approval and/or waiver of informed consent will be obtained as required by participating institutions. The TCIA cohort consists of public de-identified data and does not involve direct participant contact.

No related papers found yet.

None listed

China