Impact of SUSAN Denoising and ComBat Harmonization on Machine Learning Model Performance for Malignant Brain Neoplasms.

Abstract

Feature variability in radiomics studies due to technical and magnet strength parameters is well known and may be addressed through various pre-processing methods. However, very few studies have evaluated downstream impact of variable pre-processing on model classification performance in a multi-class setting. We sought to evaluate the impact of SUSAN denoising and ComBat harmonization on model classification performance. A total of 493 cases (410 internal and 83 external dataset) of glioblastoma (GB), intracranial metastatic disease (IMD) and primary CNS lymphoma (PCNSL) underwent semi-automated 3D-segmentation post baseline image processing (BIP) consisting of resampling, realignment, co-registration, skull stripping and image normalization. Post BIP, two sets were generated, one with and another without SUSAN denoising (SD). Radiomics features were extracted from both datasets and batch corrected to produce four datasets: (a) BIP, (b) BIP with SD, (c) BIP with ComBat and (d) BIP with both SD and ComBat harmonization. Performance was then summarized for models using a combination of six feature selection techniques and six machine learning models across four mask-sequence combinations with features derived from one-three (multi-parametric) MRI sequences. Most top performing models on the external test set used BIP+SD derived features. Overall, use of SD and ComBat harmonization led to a slight but generally consistent improvement in model performance on the external test set. The use of image pre-processing steps such as SD and ComBat harmonization may be more useful in a multiinstitutional setting and improve model generalizability. Models derived from only T1-CE images showed comparable performance to models derived from multiparametric MRI.