Delta Radiomics Texture Analysis Using Daily Setup Images In MR-Guided Pancreatic Radiotherapy
Abstract
Purpose
Magnetic resonance-guided set up images (MRg-images) acquired prior to daily treatment provide longitudinal image series that could contain treatment response information. The purpose of this work is to compare delta radiomics texture features (DF) extracted from low field strength images acquired prior to treatment with texture feature maps (TF-map) of the images in Pancreatic Ductal Adenocarcinoma (PDAC) patients.
Methods
Fraction 1 and 2 MRg-images of ten PDAC patients with the gross tumor volume (GTV) delineated daily were used for analysis. Image dynamic ranges were limited using the Collewet method and histogram quantization to re-bin images to 64 intensity levels. Gray-level run length-based gray-level nonuniformity (GLN) and run length nonuniformity (RLN) have been indicated as possibly predictive in prior delta radiomic studies. GLN and RLN were calculated for all GTVs and DF by normalizing the difference between fraction 1 and fraction 2 by fraction 1 (GLN and RLN). TF-maps were produced by calculating features of small sliding regions of interest, 5x5x5 pixels, and recording texture feature values to the center voxel. Map-based texture features were calculated the same way using the mean values of the GTVs (GLN_map and RLN_map).
Results
Half of the patients were classed as responders and half non-responders based on tumor pathology or follow up imaging. GLN for was 0.0113 and 0.0329 and RLN was 0.0082 and 0.0161 for responders and non-responders, respectively. GLN_map was 0.1736 and 0.0294 and RLN_map was -0.0263 and 0.0023 for responders and non-responders, respectively.
Conclusion
The map-based features behave differently than traditionally calculated DFs. The ability to visualize the features on the produced maps preserves spatial information and in the small cohort behaves in similar ways by quantifying image changes between fractions.