Abstract B023: Deep learning-based prediction of immune checkpoint inhibitor efficacy in brain metastases using brain MRI

Abstract Introduction: Brain metastases (BM) are an emerging challenge in modern oncology due to increasing incidence and limited treatments. Recent work by our group and others has illustrated that immune checkpoint inhibitors (ICI) are a promising therapy for treatment-refractory BM of diverse histologies. Yet, likelihood of response to ICI varies highly between patients, and the therapy often causes severe adverse effects. To build upon our findings and integrate ICI into precision medicine strategies for BM, scalable tools that quantify likelihood of response are needed. Methods: We curated a multi-institutional dataset of longitudinal mpMRI for 860 BM patients treated with ICI, assessing 2,542 BM responses based on pre- and 6-month post-treatment MRIs using standardized RANO criteria. A convolutional neural network (CNN) trained on pre-treatment mpMRI sequences predicted 6-month ICI efficacy. To ensure class balance and clinical relevance, four RANO classes were converted into a binary model: intracranial benefit including complete response (CR), partial response (PR) and stable disease (SD) vs. progressive disease (PD). Our custom CNN consists of three convolutional and two fully connected layers, with an 80/20 train-validation split. We also developed a "foundation model" for brain metastases using self-supervised contrastive learning with the SimCLR framework on studies from 9,408 patients from multiple private and public datasets. Our model has a 3D ResNet50 architecture and was pretrained on the axial T1-weighted, contrast-enhanced sequences of 11,659 BM mpMRIs. Image preprocessing was standardized for all our models and included reorientation, isotropic resampling to a 1 mm resolution, registration, N4 bias field correction, skull stripping, and normalization of image intensities to a zero mean and unit variance. The contrastive loss was computed by contrasting voxels centered around the centroid of the segmented tumor region against randomly selected non-tumor regions of the MRI studies. The model was then fine-tuned for ICI efficacy prediction using the same dataset that trained the CNN, held out during pretraining. A logistic regression model was trained on features extracted from the foundation model. Results: Our CNN achieved an area under receiver operating characteristics (AUROC) score of 0.650 on the validation set. We are actively optimizing the models, developing multi-class response models (e.g., CR/PR vs. SD vs. PD) and conducting subgroup analyses, such as histology-specific performance. Notably, fine-tuning the foundation model with a linear classifier resulted in an improved AUROC score of 0.7601. This demonstrates the potential of our pretraining approach for enhanced predictive performance. Conclusion: Our study presents one of the first Deep Learning-based efforts to predict ICI efficacy for BM. There is emerging promise in using Deep Learning to identify under-appreciated or previously unknown imaging patterns of biological significance within clinically acquired imaging. Citation Format: Melisa S. Guelen, Tobias R. Bodenmann, Mason C. Cleveland, Jay B. Patel, Felix J. Dorfner, Nelson Gil, Shreyas B. Brahmavar, Dagoberto Pulido-Arias, Jayashree Kalpathy-Cramer, Jean-Philippe Tiran, Bruce R. Rosen, Elizabeth Gerstner, Jawed Nawabi, David Wasilewski, Andrea Dell'Orco, Albert E. Kim, Christopher P. Bridge. Deep learning-based prediction of immune checkpoint inhibitor efficacy in brain metastases using brain MRI [abstract]. In: Proceedings of the AACR Special Conference in Cancer Research: Artificial Intelligence and Machine Learning; 2025 Jul 10-12; Montreal, QC, Canada. Philadelphia (PA): AACR; Clin Cancer Res 2025;31(13_Suppl):Abstract nr B023.

Fetched live from OpenAlex and de-inverted. Abstracts are not stored in this database: the inverted indexes are 8.6 GB of the frame’s 9.3 GB of text, and the host has 13 GB free.