Room: AAPM ePoster Library
Purpose: feature values are affected by imaging acquisition parameters. Harmonization can realign features making imaging data from multiple protocols comparable. This work evaluates the effect of ComBat harmonized radiomic features on predictive models for post-SBRT lung changes.
Methods: 66 repeatable and non-volume confounded radiomic features were extracted from consolidated (C) (-250 to 150HU) and ground glass (GG) (-500 to -250HU) radiographic lung changes on 3-, 6-, and 12-month follow-up images for 135 lung SBRT patients. These were combined with clinical factors to create 4 datasets: non-harmonized single (NHSP- 63 training/8 testing) and multi-protocol (NHMP- 11 different imaging protocols), ComBat harmonized with (HM) and without (HnM) protecting clinical features (121 training/14 testing). Linear mixed effects models predicting volume of radiographic changes were selected using 11- (NHMP, HM, HnM) or 9-fold (NHSP) cross validation. A direct comparison was performed by selecting the “best” (lowest k-fold information criteria) NHSP model and applying the same features to the other datasets.
Results: The C and GG changes had an average volume of 16.0mL and 26.9mL. The root mean square error (RMSE - in mL) for the “best” models in training/testing were: NHSP: 15.3/22.8 (GG) and 9.1/11.8 (C), NHMP: 68.8/22.9 (GG) and 27.9/17.7 (C), HnM: 22.6/18.9 (GG) and 13.9/25.7 (C), and HM: 25.2/15.4 (GG) and 16.2/8.4 (C). The direct comparison RMSE was: NHSP: 15.3/22.8 (GG) and 9.1/11.8 (C), NHMP: 23.6/14.0 (GG) and 17.5/8.7 (C), HnM: 23.1/20.1 (GG) and 23.1/19.2 (C), and HM: 21.6/12.5 (GG) and 16.8/8.3 (C).
Conclusion: ComBat harmonization may not improve results, if features are selected with NHSP data. When this is not possible, harmonizing the data appears to improve accuracy over NHMP, but is not as accurate as NHSP features. There is little difference between HnM and HM models unless features are selected with NHSP.
Funding Support, Disclosures, and Conflict of Interest: This work was supported in part by Varian Medical Systems (Varian). RNM, GDH, and EW: research grant support from Varian. EW and NM: research grant support from NIH. EW: royalties from Up-to-Date. GDH: research grant support from Viewray, and Siemens and honorarium from Varian. MG and NK: no disclosures.
Feature Selection, Texture Analysis, Statistical Analysis
IM/TH- Image Analysis Skills (broad expertise across imaging modalities): Feature extraction, texture analysis, radiomics