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

Abstract

Background and purpose: Feature variability in radiomics studies due to technical and magnet strength parameters is well-known and may be addressed through various preprocessing methods. However, very few studies have evaluated the downstream impact of variable preprocessing on model classification performance in a multiclass setting. We sought to evaluate the impact of Smallest Univalue Segment Assimilating Nucleus (SUSAN) denoising and Combining Batches harmonization on model classification performance.

Materials and methods: A total of 493 cases (410 internal and 83 external data sets) of glioblastoma, intracranial metastatic disease, and primary CNS lymphoma underwent semiautomated 3D-segmentation post-baseline image processing (BIP) consisting of resampling, realignment, coregistration, skull-stripping, and image normalization. Post-BIP, 2 sets were generated, one with and another without SUSAN denoising. Radiomics features were extracted from both data sets and batch-corrected to produce 4 data sets: (a) BIP, (b) BIP with SUSAN denoising, (c) BIP with Combining Batches, and (d) BIP with both SUSAN denoising and Combining Batches harmonization. Performance was then summarized for models using a combination of 6 feature-selection techniques and 6 machine learning models across 4 mask-sequence combinations with features derived from 1 to 3 (multiparametric) MRI sequences.

Results: Most top-performing models on the external test set used BIP+SUSAN denoising-derived features. Overall, the use of SUSAN denoising and Combining Batches harmonization led to a slight but generally consistent improvement in model performance on the external test set.

Conclusions: The use of image-preprocessing steps such as SUSAN denoising and Combining Batches harmonization may be more useful in a multi-institutional setting to improve model generalizability. Models derived from only T1 contrast-enhanced images showed comparable performance to models derived from multiparametric MRI.

FIG 1.

Schematic depicting the study workflow.

FIG 2.

Violin plots for all 4 feature sets using the external data show the range of mAUC across different pipelines. Feature set 1: CE_ET and F_PTR; 2: CE_ET and T2_PTR; 3: CE_ET, A_ET and F_PTR; 4: CE_ET only. A indicates ADC; F, FLAIR.

FIG 3.

Maximum mAUC heatmaps for the internal and external data for the various models based on the ML algorithm. svmRBF indicates support vector machine-Gaussian kernel; XGB, extreme gradient boosting; RF, random forest; GBRM, generalized boosted regression mode; ENET, multinomial elastic net; MAX, maximum; svmPoly, support vector machine-polynomial kernel.

FIG 4.

Histogram plot showing mean differences between the internal and external data model performance for the different pipelines. The red line depicts the mean difference in model performance.