Making Radiomics More Reproducible across Scanner and Imaging Protocol Variations: A Review of Harmonization Methods

Abstract

Radiomics converts medical images into mineable data via a high-throughput extraction of quantitative features used for clinical decision support. However, these radiomic features are susceptible to variation across scanners, acquisition protocols, and reconstruction settings. Various investigations have assessed the reproducibility and validation of radiomic features across these discrepancies. In this narrative review, we combine systematic keyword searches with prior domain knowledge to discuss various harmonization solutions to make the radiomic features more reproducible across various scanners and protocol settings. Different harmonization solutions are discussed and divided into two main categories: image domain and feature domain. The image domain category comprises methods such as the standardization of image acquisition, post-processing of raw sensor-level image data, data augmentation techniques, and style transfer. The feature domain category consists of methods such as the identification of reproducible features and normalization techniques such as statistical normalization, intensity harmonization, ComBat and its derivatives, and normalization using deep learning. We also reflect upon the importance of deep learning solutions for addressing variability across multi-centric radiomic studies especially using generative adversarial networks (GANs), neural style transfer (NST) techniques, or a combination of both. We cover a broader range of methods especially GANs and NST methods in more detail than previous reviews.

Keywords: deep learning; feature reproducibility; harmonization; medical imaging; radiomics.

Figure 1

PET and CT slices obtained from two different centers (Center 1 = Centre Hospitalier Universitaire de Sherbrooke, Sherbrooke, Canada and Center 2 = Hôpital Maisonneuve-Rosemont, Montréal, Canada). The top row shows CT images while the bottom row shows PET images. The four columns indicate four different patients. Adapted from [23].

Figure 2

Overview of harmonization methods at different stages of medical imaging.

Figure 3

Basic GAN-CIRCLE network. Here, X is a set of LR CT scans and Y is the corresponding HR CT scans. The network has two GAN modules, a low-to-high image reconstructor (Generator G, Discriminator DY) and a high-to-low image reconstructor (Generator F, Discriminator DX). Different loss functions are harmoniously coupled for training the network and monitored with regularized cycle-consistency and identity loss to prevent overfitting. Figure is adapted from [63].

Figure 4

Normalization results reused from [87] with original copy obtained from authors. The row represents a sagittal view of the scans containing the ROI nodule. Column (a) represents the target image; Column (b,e,h) shows the input images with different slice thicknesses and dosage; Columns (c,f,i) show the CNN results and Columns (d,g,j) show the GAN-based results.

Figure 5

Illustration of concept of neural style transfer using original work [101].

Figure 6

PET to CT translation using MedGAN. Figure adapted from [116] with permission from Elsevier.

Figure 7

Probability density function of homogeneity before and after applying ComBat for realignment between different CT reconstruction algorithms, reconstruction kernels and slice thicknesses. FBP: filtered back-projection. Figure reproduced from [170]. Figure reproduced with copyright permission from The Radiological Society of North America.

Figure 8

An overview of the proposed normalization method by [21] using a CT phantom. Figure adapted from [21].