Spatially Aware Acquisition-Independent Deep Learning for IVIM MRI Parameter Estimation in Patients With Esophageal Cancer.

Purpose

Neural controlled differential equations (NCDEs) recently emerged as a robust deep learning approach for quantitative MRI parameter estimation. NCDEs offer flexibility to changes in acquisition protocols by modeling the signal evolution dynamics. However, NCDE implementations operate on a voxel-by-voxel basis and cannot exploit spatial information, limiting effectiveness. The purpose of this study is to develop and validate an acquisition-independent and spatially aware neural network for intra-voxel incoherent motion (IVIM) MRI parameter estimation.

Methods

Spatially aware NCDEs (Spatial NCDEs) were evaluated in simulations across a range of acquisition protocols and signal-to-noise ratios. Performance was compared with least squares (LSQ) fitting, segmented fitting, voxel-wise NCDEs, and spatially aware neural networks (UNet). In patients with esophageal cancer, discriminative ability for predicting response to neoadjuvant chemoradiotherapy was examined.

Results

Spatial NCDEs achieved lower mean squared error (MSE) for estimating IVIM parameters than LSQ, segmented fitting, 1D NCDE, and UNet. At SNR 20, MSE was 81%, 85%, and 83% lower than LSQ, 85%, 89%, and 88% lower than segmented fitting 62%, 71%, and 52% lower than 1D NCDE and 55%, 67%, and 44% lower than UNet for D , D * , and f , respectively. In patients with esophageal cancer, Spatial NCDE-based parameter estimates showed improved, though not statistically significant, discriminative ability for predicting response to neoadjuvant chemoradiotherapy compared to LSQ.

Conclusion

Spatial NCDEs provide a robust, accessible solution for high-quality IVIM MRI parameter estimation, enabling broader adoption of deep learning-based parameter estimation.