(Thursday, 7/18/2019) 11:00 AM - 12:00 PM
Room: Stars at Night Ballroom 2-3
Purpose: To determine whether management of imaging protocols can improve radiomics prediction accuracy as variations can significantly hinder the translation of radiomics models to large-scale clinical applications.
Methods: (1) Phantom studies: CT and PET texture phantoms were scanned using multiple imaging protocols on multiple scanners. We determined the impact of imaging protocol parameters and inter-scanner variations on radiomics features using linear mixed-effects models and ICCs. Then we investigated a harmonized imaging protocol (e.g., matched kernel across manufacturers) to reduce variability. (2) Head-and-neck patient studies: We investigated the impact of a harmonized imaging protocol on radiomics outcome models. Cox proportional hazards models were created for overall survival with 726 CT and 686 PET patient images from multiple institutions, split into training/testing. This process was then repeated on a subset of the patients with matched imaging protocols.
Results: (1) A harmonized CT protocol was established; testing across 100 scanners showed that this protocol reduced the inter-scanner variability of radiomics features by 52%. For PET scans, intra-scanner reproducibility was high (average ICC=0.90), but inter-scanner reproducibility was poor and could not be managed by harmonizing the protocol. (2) The CT-based model with all patients had an AUC of 0.72, but the model for the subset of patients with matched imaging protocols did not improve this (AUC=0.55). The PET-based model had an AUC of 0.59, but the model for the subset of patients with matched imaging protocols had no covariates selected.
Conclusion: In order to reduce variability in radiomics studies, we recommend using a harmonized imaging protocol for CT and a PET texture phantom to individually assess PET scanners. However, we were unable to demonstrate a beneficial impact on survival analysis in large cohorts of head-and-neck patients. We attribute this to non-imaging protocol-based sources of variability, which may prevent translation of radiomics models between institutions.
Keywords
Taxonomy
IM/TH- Image Analysis (Single modality or Multi-modality): Imaging biomarkers and radiomics
Contact Email
