MENU

Click here to

×

Are you sure ?

Yes, do it No, cancel

Machine Learning of Tumor Cluster Dosi-Radiomics to Predict Regional Changes On Early-Response FDG PET/CT Imaging of FLARE-RT Protocol Patients

C Duan1,2,4 , W Chaovalitwongse2 , K Puk3 , P Thammasorn2 , S Wang3 , D Hippe4 , L Pierce4 , X Liu2 , J You1 , R Miyaoka4 , H Vesselle4 , P Kinahan4 , R Rengan4 , J Zeng4 , S Bowen4*, (1) Tongji University School of Economics & Management, Yangpu District, Shanghai, (2) University of Arkansas, Fayetteville, AR, (3) University of Texas, Austin, TX, (4) University of Washington School of Medicine, Seattle, WA

Presentations

(Monday, 7/30/2018) 4:30 PM - 5:30 PM

Room: Exhibit Hall | Forum 1

Purpose: Identification of high-risk tumor subregions and prediction of inhomogeneous response to radiation therapy (RT) on longitudinal imaging can guide biologically adaptive treatment strategies. We evaluated machine learning models of tumor cluster dosi-radiomics to predict regional response on early-RT FDG-PET/CT imaging.

Methods: Pre-RT (PETpre) and 3-week imaging (PETmid) during chemo-radiotherapy was performed in 19 stage IIB-IIIB non-small cell lung cancer patients enrolled on the FLARE-RT trial. Metabolic tumor volumes (MTV) were segmented on PETpre and PETmid using gradient search. K-means and hierarchical MTV voxel clusters were defined by Euclidean distance of SUV and 3D-position. Cluster number was selected to maximize CH-index (inter-cluster/intra-cluster variance ratio). PETpre and RT-dose radiomics consisted of 41 intensity and textural features (3D-64bin GLCM, GLNDM, GLSZM) within MTV and MTVclusters. Feature dimensionality was reduced by principal component analysis. Multiple linear (MLR), support vector (SVR), and decision tree (DTR) regression algorithms were employed to predict PETmid changes (Δ3wkVol, Δ3wkSUVmean). Leave-one-out cross-validated (LOOCV) root-mean-squared errors (RMSE) were calculated for MTVtumor and MTVcluster models using PETpre dosi-radiomics or benchmark features (preVol, preSUVmean, dosemean). Prediction performance was compared by Friedman ANOVA.

Results: K-means clustering (MTVhi-cluster, MTVlo-cluster) with buffer region separation produced the highest CH indices for PETpre and PETmid (p=10-15). MTVhi-cluster MLR and SVR models had improved LOOCV prediction of Δ3wkSUVmean (median RMSE=16%) compared to MTVtumor (RMSE=20%) and MTVlo-cluster (RMSE=24%, p=0.02), including lower errors using benchmark features (p=0.05). Errors in predicted Δ3wkVol trended lower for MTVtumor models (RMSE=25%) compared to MTVlo-cluster (RMSE=35%) and MTVhi-cluster (RMSE=34%, p=0.08). MLR models had smaller range of prediction errors (14-34%) relative to SVR (15-46%) and DTR (19-42%) models.

Conclusion: Prediction of tumor subregion volume and uptake changes on FDG-PET/CT during radiotherapy provides enhanced decision support for multi-resolution response assessment. Tumor cluster prediction models can be extended temporally to post-treatment response patterns and refine individualized risk-adaptive therapies.

Funding Support, Disclosures, and Conflict of Interest: This work was supported by NIH/NCI R01CA204301, the Natural Science Foundation of China (71701153), and the International Postdoctoral Exchange Fellowship program (20160087).

Keywords

PET, Lung, Texture Analysis

Taxonomy

IM- PET : Machine learning, computer vision

Contact Email