Fig 1.
Annotated ROIs in the liver area of a 3D anthropomorphic radiopaque phantom.
(a) Presents an axial view of annotated ROIs in the liver region, with green representing benign cysts, blue representing hemangioma, red representing normal liver tissue, and yellow representing liver metastasis from colon carcinoma. (b) Presenting the coronal view of the phantom. (c) Depicting a 3D rendering of the annotated ROIs.
Table 1.
Overview of CT reconstruction parameter variations.
Fig 2.
GAN workflow for image level harmonization.
The GAN’s generator learns the mapping between the target and source domains, aided by a WGAN-GP-based critic. This process generates harmonized images closely resembling the target domain while preserving source image characteristics.
Fig 3.
Architecture of the proposed shallow CNN.
Fig 4.
This figure presents an overview of the experimental setup conducted in three sub-experiments for harmonizing radiomic data. Sub-experiment 1 involves image-level harmonization using GANs in a pairwise fashion. Sub-experiment 2 focuses on feature-level harmonization through ComBat on radiomic features extracted from ROIs. Sub-experiment 3 combines both image and feature-level harmonization, employing GANs followed by ComBat to achieve comprehensive harmonization.
Fig 5.
UMAP plots with each subplot visualizing 90 samples (30 scans, for one ROI and one group), described by 93 radiomic features, across harmonization methods.
Table 2.
CCC calculations averaged across features, ROIs, and groups for each harmonization method.
Table 3.
ROI-Specific paired stability analysis results.
Table 4.
ROI-Specific classification scores for all harmonization methods averaged over all groups.
Fig 6.
Example of GAN harmonization for Group 2.
Table 5.
Image quality scores from GAN harmonization.