P Ghasemi Saghand*, I El Naqa, S Rosenburg, J Bryant, K Latifi, J Frakes, S Hoffe, E Moros, Moffitt Cancer Center, Tampa, FL

• 
  P Ghasemi Saghand

• 
  
• 
  
• 
  
• 
  
• 
  
• 
  
• 
  

PO-GePV-M-39 (Sunday, 7/10/2022)   [Eastern Time (GMT-4)]

ePoster Forums

Purpose: To test the power of tumor’s morphological changes during treatment with MR-guided Radiotherapy (MRgRT) for progression prediction using deep learning and delta-radiomics.

Methods: Twenty-four adrenal cancer and twenty-three lung cancer patients treated with 0.35T MRIdian MRgRT from 2019-2021 with median prescriptions of 50 Gy in 5 fractions were included. Sixteen adrenal and eleven lung cancer patients experienced progression with median times of 4 and 6 months. Median follow-up was 5 and 10 months for adrenal and lung cases. Each patient had 6 MRI scans, one at simulation 14-21 days before treatment, and one immediately before each fraction. GTV was segmented in the simulation and co-registered in other scans. Progression was defined as occurrence of local/distant failure before the latest follow-up.A Deep Neural Network (DNN) was developed for each cancer site using 3D ResNet blocks followed by blocks of Transformers for imaging and temporal feature extraction. Batch normalization, layer normalization, dropout, and spectral decoupling were used to improve robustness. The quality of predictions was assessed using Area under the ROC Curve (AUC) and 100 iterations of Bootstrap .632+ method.For comparison, 73 texture radiomic features were extracted and 6 classification models were trained on delta-radiomics data, i.e., feature ratios between the first and last fractions.

Results: The DNN resulted in average AUC of 0.72 (95% CI, 0.69-0.75) for adrenal and 0.74 (95% CI, 0.71-0.77) for lung cases. For comparison, Gaussian Naive Bayes and Random Forests were the best models for delta-radiomics data of adrenal and lung cases, respectively, with AUC of 0.67 (95% CI, 0.64-0.70) and 0.68 (95% CI, 0.65-0.71), with the top delta-radiomic feature being GLSZM gray level non-uniformity.

Conclusion: Our analysis demonstrated that DNN trained directly on the longitudinal 3D GTV images outperformed models trained with conventional delta-radiomics. Further analysis using external validation data are in progress.

Keywords

Not Applicable / None Entered.

Taxonomy

Not Applicable / None Entered.

Contact Email