Effect of CT Reconstruction Algorithm on the Diagnostic Performance of Radiomics Models: A Task-Based Approach for Pulmonary Subsolid Nodules.

Abstract

We investigated whether the diagnostic performance of machine learning-based radiomics models for the discrimination of invasive pulmonary adenocarcinomas (IPAs) among subsolid nodules (SSNs) was affected by the proportion of images reconstructed with filtered back projection (FBP) and model-based iterative reconstruction (MBIR) in datasets used for feature extraction. This retrospective study included 60 patients (23 men and 37 women; mean age, 61.4 years) with 69 SSNs (54 part-solid and 15 pure ground-glass nodules). Preoperative CT scans were reconstructed with both FBP and MBIR. A total of 860 radiomics features were obtained from the entire nodule volume, and 70 resampled nodule datasets with an increasing proportion of nodules with MBIR-derived features (from 0/69 to 69/69) were prepared. After feature selection using neighborhood component analysis, support vector machines (SVMs) and an ensemble model were used as classifiers for the differentiation of IPAs. The diagnostic performances of all blending proportions of reconstruction algorithms were calculated and analyzed. The ROC AUC and the diagnostic accuracy of the radiomics models decreased significantly as the number of nodules with MBIR-derived features increased, and this relationship followed cubic functions (R2 = 0.993 and 0.926 for SVM; R2 = 0.993 and 0.975 for the ensemble model; p < 0.001). The magnitude of variation in AUC due to the reconstruction algorithm heterogeneity was 0.39 for SVM and 0.39 for the ensemble model. Inclusion of CT scans reconstructed with MBIR for radiomics modeling can significantly decrease diagnostic performance for the identification of IPAs.