Adaptive multi-scale attention-enhanced V-NET for accurate and efficient multi-shell diffusion MRI estimation

Abstract

Multi-shell acquisition protocols sample diffusion signals at different b -values. While higher b -values (e.g., \(b = 2000\, \mathrm {s/mm^2}\) ) offer increased sensitivity to microstructural features, their acquisition is time-intensive and suffers from reduced signal-to-noise ratio and motion sensitivity, which limits their routine clinical applicability. To address this challenge, we propose a deep learning framework for predicting high b -value ( \(b = 2000\, \mathrm {s/mm^2}\) ) spherical harmonic (SH) coefficients directly from low b -value ( \(b = 1000\, \mathrm {s/mm^2}\) ) SH coefficients. Our method is based on a simplified V-NET architecture augmented with adaptive multi-scale attention. This attention mechanism dynamically learns optimal receptive field combinations through scale-weighting networks. Unlike conventional fixed-scale approaches, our model incorporates three dilated convolutional branches with adaptive weighting, cross-scale feature fusion, and integrated spatial–channel attention. These components enable context-aware feature extraction that emphasizes the most relevant scales for diffusion pattern recognition. Experimental results demonstrate that the proposed method achieves accurate reconstruction with significantly reduced computational cost compared to modern state-of-the-art approaches.