Click here to


Are you sure ?

Yes, do it No, cancel

Creation of An Automated Hand-Crafted Radiomics Methodology and Assessment of Its Potential to Contribute to a Prospective Trial

E Carver*, J Snyder, D Bergman, M Shah, S Siddiqui, N Wen, Henry Ford Health System, Detroit, MI


(Sunday, 7/12/2020)   [Eastern Time (GMT-4)]

Room: AAPM ePoster Library

Purpose: Radiomics harbors tremendous potential to improve medical treatments by enhancing clinically usable models that inform individualized approaches and/or responses to therapy, such as defining factors that contribute to cohorts of patients that benefit from prospective trials (PT). This study’s purpose was two-fold. The first aspect involved creating an automated radiomics workflow based on neural networks that could be used in the clinic without increasing clinician workload. The second aspect was to apply this workflow to identify indicative radiomic features that could potentially aid in defining patients with recurrent Glioblastoma (GBM) that would benefit from PT assessing fractionated stereotactic radiation therapy of 32Gy to the gross tumor and 24Gy to the surrounding flair region over four fractions.

Methods: Pre/Post T2-(T2W), T1-(T1W) weighted, T1-contrast enhanced(T1CE), fluid attenuated inversion recovery (FLAIR) MR brain images, clinical factors, and length of survival were obtained from 15 recursive GBM patients (9 participating, 6 control) enrolled in a prospective trial. Pre-processing included resampling, registration, normalization, and skull stripping, using my U-Net. Image biomarker standardization initiative was followed; however, extensive harmonization was not required as same MRI protocol/scanner was used for each patient. Tumor was delineated by U-Net trained with 210 pre-operative T2W, T1W, T1CE, FLAIR and synthetic MR brain images created by a Generative Adversarial Network. Cancer Imaging Phenomics Toolkit was used to extract 20 features per modality for pre and post image sets. Individual feature’s relevance was determined by showing significance in univariate Cox Proportional Hazards and highly ranked by SVM recursive feature elimination versus overall survival post-treatment

Results: patients receiving FSRT; one clinical feature and one radiomic feature from pre-treatment, delta and post-treatment datasets. No features were found for control.

Conclusion: preliminary results and ease of workflow allow for radiomic analysis to predict the overall survival of recurrent GBM patients with FSRT.

Download ePoster [PDF]


Image Analysis


IM/TH- Informatics: Informatics in Therapy (general)

Contact Email