MENU

Click here to

×

Are you sure ?

Yes, do it No, cancel

Machine Learning of MAA SPECT Lung Perfusion Radiomics to Predict Radiation and Immune-Mediated Pneumonitis in Patients with Locally Advanced Non-Small Cell Lung Cancer

H Thomas T1, J Zeng2, P Kinahan2, R Miyaoka2, H Vesselle2, R Rengan2, S Bowen2*, (1) Christian Medical College Vellore, Vellore, TN, IN, (2) University of Washington, School of Medicine, Seattle, WA

Presentations

(Sunday, 7/12/2020)   [Eastern Time (GMT-4)]

Room: AAPM ePoster Library

Purpose: Functional lung radiomics for pneumonitis risk stratification in the setting of chemoradiation and immunotherapy is untested. Machine learning models to predict incidence of combined radiation and immune-mediated pneumonitis were generated from lung perfusion radiomics under robust perturbation of feature extraction and feature selection methods.

Methods: Thirty patients with locally advanced non-small cell lung cancer (NSCLC) enrolled on the FLARE-RT trial (NCT02773238) and received chemoradiotherapy with functional lung avoidance planning and consolidative durvalumab immunotherapy. Within tumor-subtracted lung regions, PyRadiomics features (110 shape/intensity/texture) were extracted on pre-treatment and 3-month post-treatment MAA SPECT/CT perfusion images using fixed bin size (FBS=64bins) and fixed bin width (FBW=25CNTS) discretization. Two feature selection methods were tested for model stability: M1 included (i) inter-patient variance inflation, (ii) pre/post-treatment variance inflation, (iii) co-linearity reduction, (iv) 60 bootstrap iterations of least absolute shrinkage and selection operator (LASSO); M2 applied 60 LASSO bootstraps only. LASSO logistic regression of the top 5 pre-treatment radiomic features was conducted over 100 stratified random samples of 80% training / 20% testing datasets. Ensemble performance of testing datasets for predicting combined radiation and immune-mediated CTCAEv4 Grade 2+ pneumonitis (40% incidence rate) was quantified by the area-under-ROC-curve (AUC).

Results: ZoneSizeGrayLevelNonUniformity (GLSZM-GLN) was selected in 75-84% of bootstraps by LASSO for predicting pneumonitis across all models. FBW-M1 model with GLSZM-GLN achieved the highest testing performance (AUC=0.705, OR=2.09, p=0.05), followed by FBW-M2 with GLSZM-GLN (AUC=0.692, OR=2.09, p=0.05). FBW generated single-feature parsimonious models that outperformed multivariate FBS models (AUC=0.589-0.672, OR=1.32-1.97, p<0.57).

Conclusion: The GLSZM-GLN texture feature of lung perfusion heterogeneity was identified as a potential imaging biomarker of combined radiation and immune-mediated pneumonitis risk. Prediction model performance was more sensitive to gray-level discretization than feature selection method. This functional lung imaging biomarker will be tested in larger patient populations as a complement to other known risk factors of pneumonitis.

Funding Support, Disclosures, and Conflict of Interest: Research supported by NIH/NCI R01CA204301.

Keywords

SPECT, Perfusion Imaging

Taxonomy

IM- SPECT : Radiomics

Contact Email