LAUSR

Editorial for “An Unsupervised Deep Learning Approach for Dynamic-Exponential Intravoxel Incoherent Motion MRI Modeling and Parameter Estimation in the Liver” Intravoxel incoherent motion (IVIM) magnetic resonance imaging (MRI) has gained renewed momentum recently. While it was initially developed for imaging of the brain, it is increasingly applied to abdominal imaging, including assessment of renal fibrosis, evaluation of liver tumors, or staging of liver cirrhosis. IVIM allows simultaneous and separate quantification of the diffusion of free water molecules with Gaussian distribution (true diffusion) and microcirculation in the randomly oriented capillaries (pseudo-diffusion). Both true diffusion and pseudo-diffusion contribute to the signal intensity in diffusion-weighted MRI. Importantly, although pseudodiffusion and noise affect signals acquired at high b values, IVIM effects due to microscopic perfusion must be considered at low b values. The distribution of low and high b values can therefore affect the metrics of the IVIM parameters. Using images acquired at increasing b-values, IVIM enables a separation of the relative contribution of true diffusion and pseudo-diffusion, and thus a noninvasive quantification of tissue perfusion without the need for intravenous contrast agents. Methods for calculating IVIM have been continuously refined over the years. Still, current IVIM models cannot reliably distinguish between healthy liver tissue and fibrosis or between benign and malignant lesions, so they were extended to multiexponential curve fitting approaches. From the original bi-exponential models as described by Le Bihan et al, tri-exponential models at small b-values emerged for organs with dual blood supply such as the liver. Currently, curve fitting approaches with a combination of least squares (LSQ) and Akaike information criterion (AIC) provide the most accurate estimation of IVIM. However, IVIM models are less stable than monoexponential diffusion models, require the fitting of more variables, and are susceptible to low signal-to-noise ratios, respiration, or cardiac motion artifacts. Moreover, they are computationally intensive when applied at the voxel level and need to be recalculated for each new examination, which is associated with high computational requirements and consequently long inference times. All of this limits clinical transferability. According to a recent survey, IVIM is used in clinical practice in only a small proportion of institutions. Therefore, further technical innovations are needed for clinical implementation of IVIM-MRI, such as protocols with shorter acquisition times or novel, fast, and robust methods to calculate perfusion in IVIM. One promising approach is the application of convolutional neural networks, especially convolutional autoencoders. Advantages of autoencoder networks are a robustness to image noise and high computational efficiency, even when calculations are performed on the voxel level. While training of neural networks is time consuming and requires high-performance hardware, deployment of a trained network is very fast and has low hardware requirements. Autoencoders consists of two parts: an encoder and decoder. During training, the input image is first passed through the encoder where the image is mapped into a highdimensional, efficient numeric representation. The decoder then attempts to recreate the input from the numeric representation and learns to ignore insignificant data, such as image noise. However, one can also train the autoencoder to produce different results than the input image. In this issue of JMRI, an article by Zhou et al proposes an unsupervised deep learning approach to model IVIM in the liver, where they used a convolutional autoencoder to create parametric maps of IVIM from diffusion-weighted signals. Based on 10 patients, this innovative proof-of-concept study shows that the use of deep neural networks can provide significantly better IVIM parameter estimation and improved parameter maps, while being faster and more computationally efficient than the conventional LSQ–AIC approach. In