Click here to


Are you sure ?

Yes, do it No, cancel

Standardizing Patient Orientation to Improve Generalization of Radiomics Models

A Iyer*, J Oh , M Thor , J Deasy , A Apte , Memorial Sloan Kettering Cancer Center, New York, NY


(Tuesday, 7/16/2019) 9:30 AM - 10:00 AM

Room: Exhibit Hall | Forum 6

Purpose: Radiomics feature calculation uses directional information, which is influenced by patient orientation. We propose a standardization approach to correct for patient orientation prior to feature extraction for improved generalization of radiomics models.

Methods: The proposed approach was evaluated using a radiomics-based Random Forest (RF) classifier trained to detect slices with dental artifacts in head and neck (HN) CT scans. A dataset of 44 CT scans (1165 axial slices) from our institution were used to train the classifier using 26 GLCM-based features extracted using open-source software CERR. To standardize patient positioning, training CTs were rigidly registered to a reference scan using open-source tool Plastimatch. This resulted in consistent coordinate system and voxel resolutions across the training cohort. To preserve texture characteristics of the original scans, 6-degree-of-freedom rigid registration was employed, which corrects for differences in patient positioning due to translations and rotations. Performance was compared against the conventional approach in which voxel resolution alone is standardized using sinc-resampling.

Results: RF classifiers were built on 100 bootstrapped samples using conventional (C1) and proposed (C2) approaches. The resulting models were evaluated on (1) internal hold-out samples (excluded from the bootstrap) and (2) external dataset from The Cancer Imaging Archive (TCIA), consisting of 24 HN CT scans (679 slices) showing greater variation in patient positioning compared to the internal training dataset. C1 and C2 showed similar performance (AUC_C1=0.96, AUC_C2=0.95) on the internal hold-out data. However, while C2 performed similarly well on the external TCIA dataset (AUC_C2=0.97), C1 resulted in a decrease in performance (AUC_C1=0.89).

Conclusion: This work demonstrates the potential for improved generalization of radiomics models through standardization of patient positioning prior to feature extraction.

Funding Support, Disclosures, and Conflict of Interest: This research was partially funded by NIH grant 1R01CA198121 and NIH/NCI Cancer Center Support grant P30 CA008748.


Texture Analysis, Feature Extraction


IM/TH- Image Analysis (Single modality or Multi-modality): Imaging biomarkers and radiomics

Contact Email