MENU

Click here to

×

Are you sure ?

Yes, do it No, cancel

Comparison in Classification Performance of Radiation Pneumonitis Between Two Delta Radiomics Logistic Regression Models

J Foy*, H Al-Hallaq , S Armato , The University of Chicago, Chicago, IL

Presentations

(Sunday, 7/14/2019) 1:00 PM - 2:00 PM

Room: Stars at Night Ballroom 2-3

Purpose: Previous studies have shown promise in classifying patients with radiation pneumonitis (RP) by averaging changes in radiomics features over regions of interest (ROIs) in each patient. This study compares this approach to an alternative method using changes in individual ROI pairs during model training.

Methods: Pre- and post-radiotherapy (RT) CT scans were acquired from 105 esophageal cancer patients. Twenty patients developed RP grade≥2. ROIs were randomly placed in the lung volume of the pre-RT scan. The vector map obtained from deformable registration anatomically matched corresponding ROIs in the post-RT scan. Eight radiomics features were calculated for each ROI, and changes in radiomics features were calculated. Two logistic regression models were constructed classifying patients with RP for each feature: one trained with differences between individual ROI pairs (n=4474), denoted M(Ind), and one trained with differences in feature values averaged over all ROIs for each patient (n=105) denoted, M(Avg). Both models were tested using differences in features averaged over all ROIs for each patient, and the area under the curve (AUC) was calculated. Sampling was repeated 1000 times resulting in a mean AUC value and 95% confidence intervals. ANOVA was used to determine significant associations between changes in feature values and RP status for each feature using both models.

Results: Among M(Avg) and M(Ind) models, 5/8 and 7/8 of the features were significantly correlated with RP status, respectively. Only minimum was not correlated for either model. AUC values ranged from 0.58-0.74 and 0.59-0.75 for M(Avg) and M(Ind), respectively, with mean AUC values from M(Ind) models being higher than mean AUC values from M(Avg) models for 6/8 features, but these differences were not significant.

Conclusion: Using averages in changes in radiomics features to train models may dampen textural changes in RP patients resulting in fewer features showing significant relationships with RP development.

Keywords

Quantitative Imaging, Image Analysis, Computer Vision

Taxonomy

IM/TH- Image Analysis (Single modality or Multi-modality): Classification methods

Contact Email