{
  "abstract": "Background Predicting intracranial progression-free survival (iPFS) is crucial for managing ALK-positive NSCLC patients with brain metastases undergoing tyrosine kinase inhibitor (TKI) therapy. Quantitative radiomics (analyzing medical images beyond visual interpretation) offers a potential non-invasive method for prognostication. This study aimed to evaluate the predictive performance of MRI-based radiomic features, assessed at baseline and first follow-up, for iPFS in this population.Methods Retrospective data from 98 ALK+ NSCLC patients with brain metastases treated with ALK TKIs (brigatinib, crizotinib) were analyzed. Radiomic features were extracted from baseline and first follow-up (~8 weeks) post-contrast T1-weighted (PCT1w) and FLAIR MRI sequences. We evaluated two feature sets: baseline only and baseline combined with delta features. Multivariate survival models were developed to predict continuous iPFS using Cox proportional hazards and classify patients into short or long iPFS categories based on 6- and 12-month thresholds. Lesion-level features were aggregated to the patient level using volume-weighted averaging across all lesions or by selecting features from the single largest lesion, when relevant. Model performance was assessed using C-index and AUC via cross-validation. An exploratory lesion-level volumetric response prediction model was also built using baseline features.Results Models utilizing only baseline radiomic features demonstrated moderate predictive capability ( figure 1). The optimal baseline model for continuous iPFS achieved a C-index of 0.68. For classification tasks, baseline models yielded AUCs of 0.75 for predicting iPFS > 6 months and 0.72 for iPFS > 12 months, often leveraging combined PCT1w and FLAIR features, particularly from the largest lesion. Incorporating delta radiomic features significantly improved predictive accuracy across all endpoints (figure 2). The top-performing continuous iPFS model using delta features reached a C-index of 0.80, primarily driven by changes in FLAIR-derived features aggregated across all lesions. Delta feature-based classification models achieved strong performance, with an AUC of 0.84 for 6-month iPFS and an AUC of 0.87 for 12-month iPFS. Key predictive delta features included changes in texture, intensity, and shape metrics derived from both sequences. The exploratory lesion-level volumetric response model showed moderate discrimination (AUC 0.70) but was impacted by dataset class imbalance (92% controlled lesions).Conclusions Longitudinal delta radiomic features, capturing dynamic changes between baseline and early follow-up MRI scans, significantly enhance the prediction of iPFS compared to using only baseline features in ALK+ NSCLC patients with brain metastases receiving TKI therapy. These findings highlight the potential of dynamic radiomic analysis, integrating multi-sequence MRI data, as a powerful non-invasive tool for prognostication and monitoring treatment response.Abstract 512 Figure 1ROC Curve or time dependent AUC and feature importance plot of models using optimal baseline radiomics featuresAbstract 512 Figure 2ROC Curve or time dependent AUC and feature importance plot of optimal models using baseline radiomics and delta radiomic features",
  "authors": [
    {
      "affiliations": [
        "Takeda Development Center Americas, Inc., Cambridge, MA, USA"
      ],
      "name": "Jayant Narang"
    },
    {
      "affiliations": [
        "Royal Marsden Hospital NHS Foundation Trust, London, UK"
      ],
      "name": "Sanjay Popat"
    },
    {
      "affiliations": [
        "Radiomics Bio, Liège, Belgium"
      ],
      "name": "Wim Vos"
    },
    {
      "affiliations": [
        "Radiomics Bio, Liège, Belgium"
      ],
      "name": "Andrea Corsi"
    },
    {
      "affiliations": [
        "Radiomics Bio, Liège, Belgium"
      ],
      "name": "Nikhil Sindhwani"
    },
    {
      "affiliations": [
        "Radiomics Bio, Liège, Belgium"
      ],
      "name": "Carlos C Meca"
    },
    {
      "affiliations": [
        "Radiomics Bio, Liège, Belgium"
      ],
      "name": "Nathan Tsoutzidis"
    },
    {
      "affiliations": [
        "Takeda Development Center Americas, Inc., Cambridge, MA, USA"
      ],
      "name": "Matt McMahon"
    },
    {
      "affiliations": [
        "Takeda Development Center Americas, Inc., Cambridge, MA, USA"
      ],
      "name": "Chien-Lin Yeh"
    },
    {
      "affiliations": [
        "Takeda Development Center Americas, Inc., Cambridge, MA, USA"
      ],
      "name": "Ronald Gounden"
    },
    {
      "affiliations": [
        "Takeda Development Center Americas, Inc., Cambridge, MA, USA"
      ],
      "name": "Christopher Danes"
    },
    {
      "affiliations": [
        "Takeda Development Center Americas, Inc., Cambridge, MA, USA"
      ],
      "name": "Ozlem Yardibi"
    }
  ],
  "title": "512 Predictive value of baseline and longitudinal MRI radiomics for intracranial progression-free survival in ALK-positive NSCLC patients with brain metastases",
  "uid": "4811d728-c3cd-57f5-8a16-2c038a9103e5"
}