D DeVries^1,2*, F Lagerwaard³, J Zindler^4,5, T Yeung⁶, G Rodrigues^1,2, G Hajdok^1,7, A Ward^1,2, (1) Western University, London, ON, CA, (2) London Regional Cancer Program, London, ON, CA, (3) Amsterdam UMC, Amsterdam, NL, (4) Haaglanden Medical Centre, Den Haag, NL, (5) Holland Proton Therapy Centre, Delft, NL, (6) Reflexion Medical Inc, (7) Niagara Health, ON, CA

• 
  D DeVries

• 
  
• 
  
• 
  
• 
  
• 
  
• 
  
• 
  

PO-GePV-M-59 (Sunday, 7/10/2022)   [Eastern Time (GMT-4)]

ePoster Forums

Purpose: Prediction of brain metastasis (BM) response to stereotactic radiosurgery (SRS) has been explored using machine learning and magnetic resonance imaging (MRI) radiomics. This study provides critical analysis of this technique with respect to BM volume and MR scanner dependencies.

Methods: A dataset of 123 BMs across 99 SRS patients was assembled including pre-treatment T1w-CE MRI, SRS outcomes (progression versus non-progression), and 12 non-imaging clinical features. A set of 107 radiomic features were extracted from the T1w-CE MRI scans after the scans were interpolated to a common resolution and Z-score normalized. A random decision forest model and 250 bootstrapped resamples of the dataset were used to predict SRS response based on the pre-treatment data. To determine BM volume dependency effects, model accuracy was assessed for BMs <7.5cc versus BMs >7.5cc. Repeat experiments removing volume-correlated radiomic features were also performed. Lastly, experiments were performed to compare the relative accuracy between the three primary scanner models represented in the study.

Results: The accuracy of the predictive model was found to be dependent on BM volume, with BMs <7.5cc being predicted more accurately than BMs >7.5cc. By removing volume-correlated features, the dependency of accuracy on volume was eliminated, while overall performance remained consistent. The MR scanner analysis revealed that one scanner model decreased predictive accuracy, suggesting that the radiomic features from this scanner were in a different data domain than the features from the other two scanners.

Conclusion: As shown, a predictive model’s performance can be dependent on BM volume, which limits the model’s clinical utility. This study introduces a technique to effectively remove this dependence. It was also found that variability between scanner models can be accounted for, but not for all scanner models. This finding is highly relevant as multi-centre validation of predictive models will likely incorporate multiple MR scanner models.

Funding Support, Disclosures, and Conflict of Interest: Study supported by Western University, Lawson Health Research Institute, the Ontario Graduate Scholarship program, and the NSERC PGS-D program. No disclosures or conflict of interest to report.

Keywords

Brain, Stereotactic Radiosurgery, CAD

Taxonomy

IM/TH- Image Analysis (Single Modality or Multi-Modality): Computer-aided decision support systems (detection, diagnosis, risk prediction, staging, treatment response assessment/monitoring, prognosis prediction)

Contact Email