E Carver*, Z Dai, J Snyder, B Griffith, L Poisson, L Rogers, N Wen, Henry Ford Health System, Detroit, MI

• 
  E Carver

• 
  
• 
  
• 
  
• 
  
• 
  
• 
  
• 
  

SU-E-TRACK 6-4 (Sunday, 7/25/2021) 3:30 PM - 4:30 PM [Eastern Time (GMT-4)]

Purpose: This study aims to expand upon published literature in radiomics and deep learning to: 1) Apply deep learning to segment GBM on MRI, generating consistent ROI for radiomics. 2) Effectively employ harmonization to find a radiomic signature set (RSS) that competitively predicts GBM patient overall survival on independent datasets. 3) Integrate RSS with clinical/genetic patient information to perform a complete prognostic assessment.

Methods: 180 MR datasets including T1, T2, T1CE, and Flair were obtained through BraTS, IVYGAP, and our institution (140/18/22) for training/ independent validation. Preprocessing included resampling, registration, skull-stripping, N4 bias-correction, and normalization. Whole Tumor, Tumor Core and Enhanced Tumor were delineated on the multi-modality MRI by physicians in the training dataset. Contours used for validation datasets were generated by 2D U-Net employing synthetic MR (synMR) data augmentation by GAN. Features were extracted by CAPTK, a radiomic toolbox, and feature harmonization included ComBat with quantile normalization. Feature selection was done by LASSO and a 12-feature signature set was constructed. Predictive power of RSS was assessed by calculating the area under the curve of a receiver operating characteristic curve (AUC-ROC) created by Support Vector Machine (SVM). RSS robustness was assessed using independent validation datasets. Complete prognostic assessment was performed by comparing RSS against known clinical/genetic prognostic features (Age, Karnofsky Performance Scale, IDH1, MGMT, EGFR3) using a sub-dataset containing 59 of 180 patients, based on availability.

Results: U-Net performance and RSS prognostic ability was assessed using independent auto-contoured datasets; AUC for IVYGAP and Institutional datasets are 0.93 and 0.96, respectively. Using the sub-dataset, RSS outperformed known clinical/genetic features; AUC of 0.57. Including RSS with clinical/genetic features improved AUC to 0.83.

Conclusion: U-Net performance shows ability to fully automate RSS workflow. RSS shows encouraging preliminary results as it shows strong predictive power on independent datasets and outperforms/aids commonly used clinical/genetic features.

Handouts

Keywords

Not Applicable / None Entered.

Taxonomy

Not Applicable / None Entered.

Contact Email