Deep Generative Medical Image Harmonization for Improving Cross‐Site Generalization in Deep Learning Predictors

Abstract

Background In the medical imaging domain, deep learning‐based methods have yet to see widespread clinical adoption, in part due to limited generalization performance across different imaging devices and acquisition protocols. The deviation between estimated brain age and biological age is an established biomarker of brain health and such models may benefit from increased cross‐site generalizability. Purpose To develop and evaluate a deep learning‐based image harmonization method to improve cross‐site generalizability of deep learning age prediction. Study Type Retrospective. Population Eight thousand eight hundred and seventy‐six subjects from six sites. Harmonization models were trained using all subjects. Age prediction models were trained using 2739 subjects from a single site and tested using the remaining 6137 subjects from various other sites. Field Strength/Sequence Brain imaging with magnetization prepared rapid acquisition with gradient echo or spoiled gradient echo sequences at 1.5 T and 3 T. Assessment StarGAN v2, was used to perform a canonical mapping from diverse datasets to a reference domain to reduce site‐based variation while preserving semantic information. Generalization performance of deep learning age prediction was evaluated using harmonized, histogram matched, and unharmonized data. Statistical Tests Mean absolute error (MAE) and Pearson correlation between estimated age and biological age quantified the performance of the age prediction model. Results Our results indicated a substantial improvement in age prediction in out‐of‐sample data, with the overall MAE improving from 15.81 (±0.21) years to 11.86 (±0.11) with histogram matching to 7.21 (±0.22) years with generative adversarial network (GAN)‐based harmonization. In the multisite case, across the 5 out‐of‐sample sites, MAE improved from 9.78 (±6.69) years to 7.74 (±3.03) years with histogram normalization to 5.32 (±4.07) years with GAN‐based harmonization. Data Conclusion While further research is needed, GAN‐based medical image harmonization appears to be a promising tool for improving cross‐site deep learning generalization. Level of Evidence 4 Technical Efficacy Stage 1

Applications of Radiomics and Radiogenomics in High-Grade Gliomas in the Era of Precision Medicine

“…Through such consortiums, multiple aspects of radiomic analysis reproducibility can be tackled, and guidelines be provided for the research community. As an example, the impact of image preprocessing techniques on reproducibility of radiomic feature computations, as regulated by the Image Biomarker Standardization Initiative (IBSI), and the potential of image harmonization techniques to mitigate the variability of MRI scans [ 94 , 95 , 96 , 97 , 98 ] can be evaluated in a single controlled setting on a diverse cohort of data.…”

Commentary to “Translational machine learning for child and adolescent psychiatry”

“…Three approaches are particularly promising: Image harmonization to a standardized reference protocol. Deep learning models have been particularly promising for this goal (Bashyam, 2021), as they leverage the ability to remove subtle differences among images while preserving anatomical or functional content. These methods are complemented by statistical harmonization methods (Pomponio, 2020), which utilize subsets of the data that can be considered comparable (e.g., subsets of healthy controls with similar demographics), and infer statistical mappings that remove undesirable heterogeneity that leads to poor generalization. Very large‐scale training in diverse databases, which ensures that the models have seen ‘everything they need to see’; these approaches are common in computer vision (Deng, 2009) and are only now beginning to become feasible using very large databases of clinical images from health systems. State‐of‐the‐art machine learning methods for domain adaptation (Ganin, 2016), which teach the models to adapt to new characteristics of the data to be classified.…”

“…Image harmonization to a standardized reference protocol. Deep learning models have been particularly promising for this goal (Bashyam, 2021), as they leverage the ability to remove subtle differences among images while preserving anatomical or functional content. These methods are complemented by statistical harmonization methods (Pomponio, 2020), which utilize subsets of the data that can be considered comparable (e.g., subsets of healthy controls with similar demographics), and infer statistical mappings that remove undesirable heterogeneity that leads to poor generalization.…”

Goal-specific brain MRI harmonization

“…Site information can then be added to the latent representations to “reconstruct” the MRI data. Another popular approach is the use of generative adversarial networks and cycle consistency constraints (Zhu et al, 2017; Zhao et al, 2019; Dewey et al, 2019; Modanwal et al, 2020; Bashyam et al, 2021).…”

“…Another popular approach is the use of generative adversarial networks and cycle consistency constraints (Zhu et al, 2017;Zhao et al, 2019;Dewey et al, 2019;Modanwal et al, 2020;Bashyam et al, 2021).…”