Click here to

Session: MRI: New Algorithms, Techniques, and Applications I [Return to Session]

Using Radiomic Study Design Recommendations to Optimize Classification Performance in Brain Metastases

D Mitchell*, S Buszek, B Tran, H Liu, C Chung, UT MD Anderson Cancer Center, Houston, TX


SU-E-207-7 (Sunday, 7/10/2022) 1:00 PM - 2:00 PM [Eastern Time (GMT-4)]

Room 207

Purpose: Having previously developed recommendations for radiomic study design to improve reproducibility in patients receiving Gamma Knife treatment for brain metastases, we now compare the impact of recommendations on classifier performance for differentiating tumor progression and necrosis.

Methods: Under IRB approval, this analysis includes a retrospective cohort of 154 patients receiving Gamma Knife for brain metastases who had contrast-enhanced T1-weighted (T1+C) MR images acquired using 2-D spin echo (SE) or 3-D spoiled gradient echo (SPGR) sequences. Study design recommendations were previously developed on a 29-patient subset with both 2-D SE and 3-D SPGR T1+C series within the same exam. These were: 1) Fixed bin count intensity discretization for all studies; for studies with 2-D and 3-D datasets, 2) prefer 2-D feature extraction from non-isotropic spatial normalization and 3) exclude high-variability features from downstream analyses. Experienced physicians contoured tumor volumes using semi-automated methods, and tumor/necrosis was determined by histology. Radiomic features were extracted using several combinations of pre-processing variables in PyRadiomics (2-D or 3-D feature extraction, 0.4297x0.4297x5-mm or 1x1x1-mm spatial normalization, 64-bin or 128-bin intensity discretization) and clinical variables obtained via chart review. Z-score feature normalization and principal component analysis feature reduction (components represented 95% explained variance) were implemented. RUSBoosted trees models were selected to account for class imbalance and trained on each feature extraction. Performance was assessed using 5-fold cross-validation.

Results: Training on features using 0.4297x0.4297x5-mm spatial normalization and 2-D feature extraction (highest recommendation for heterogeneous datasets) yielded mean 65.9% accuracy and 0.60 area under the curve (AUC). 1x1x1-mm spatial normalization and 3-D feature extraction (less strongly recommended) resulted in mean 61.7% accuracy and 0.64 AUC. 1x1x1-mm spatial normalization and 2-D feature extraction (not recommended) produced mean 59.1% accuracy and 0.49 AUC.

Conclusion: Several study design recommendations were proposed to improve reproducibility. This work suggests implementing these improves classifier performance.

Funding Support, Disclosures, and Conflict of Interest: Funding provided in part by MD Anderson - CCSG Radiation Oncology and Cancer Imaging Program Grant.


MRI, Brain, Classifier Design


IM- MRI : Radiomics

Contact Email