MENU

Click here to

×

Are you sure ?

Yes, do it No, cancel

Comparison of Single and Multi-Modality CT and MR Radiomics Models for Overall Survival and Local Control for Non-Small Cell Lung Cancer

R N Mahon1*, G D Hugo2 , E Weiss1 , (1) Virginia Commonwealth University, Richmond, Virginia, (2) Washington University School of Medicine, St. Louis, MO

Presentations

(Thursday, 8/2/2018) 7:30 AM - 9:30 AM

Room: Davidson Ballroom B

Purpose: To develop single and multi-modality MRI and CT radiomics models for lung cancer using a normal tissue control method to assist in establishing robust models.

Methods: Pre-treatment T1-weighted Volumetric Interpolation Breath-Hold Examination (VIBE), true fast MRI with steady state precession (TRUFISP), and 4D helical CT scans were acquired for 15 patients with non-small cell lung cancer on an IRB-approved protocol. Fifty-nine texture features at 5 wavelet decomposition ratios were extracted from the delineated primary tumor and unirradiated normal muscle tissue. Repeatable texture features from the unfiltered and averaged wavelet ratios were clustered respectively and the most predictive (univariate) or most correlated (medoid) feature from each cluster was selected for modeling overall survival at 12 (OS_12), 18 (OS_18), and 24 (OS_24) months, and local control (PT_endTx) from both the primary tumor and the normal tissue per modality. Finally, a primary tumor multi-modality imaging model was created.

Results: For the normal tissue control, significant models were found for only the TRUFISP (3) and VIBE (4) images predominantly utilizing the univariate selection method (6/7) with only 1 significant medoid model. Significant primary tumor models were found for OS_12 (all modalities), OS_18 (TRUFISP and VIBE) and OS_24 (VIBE). There was no overlap of features between the primary tumor and control. Minimal difference in the accuracy and number of significant models was found for the VIBE and CT images using unfiltered versus the wavelet averaged features while the TRUFISP appeared to favor the unfiltered features. Accuracy of the top 3 single modality models (72%-81%) was comparable to the top 3 multi-modal models (67%-85%), with TRUFISP features appearing in 26/36 models.

Conclusion: Medoid feature selection may reduce spurious results while the wavelet decomposition is less important. Exploration of multimodality MR and CT models using these model structures is warranted.

Funding Support, Disclosures, and Conflict of Interest: GDH and EW receive research support from Varian Medical Systems, and have grants from the NIH. EW receives royalties from UpToDate. RNM has no conflicts.

Keywords

Lung, MRI, CT

Taxonomy

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

Contact Email