Reproduce “Sparse Reconstruction of Optical Doppler Tomography with Alternative State Space Model and Attention” on OpenResearch

About this paper

Zhenghong Li, Jiaxiang Ren, Wensheng Cheng, Yanzuo Liu, Congwu Du, Yingtian Pan, Haibin Ling

Researchers at Stony Brook University developed ASSAN, a deep learning framework for sparse Optical Doppler Tomography (ODT) reconstruction, which leverages Mamba State Space Models for depth-correlated information and gated self-attention for inter-A-scan features. This approach consistently outperforms other state-of-the-art methods in PSNR and SSIM on mouse cortex datasets, enabling high-fidelity ODT imaging from significantly fewer raw scans.

Key insights

• Optimal ODT reconstruction benefits from specialized, axis-wise processing that accounts for the anisotropic nature of ODT B-scans, particularly the distinct characteristics of A-line (depth) and B-line (width) information.
• Mamba State Space Models (SSMs) are effective for capturing long-range dependencies in the depth dimension of A-lines, offering computational efficiency superior to 2D selective scan approaches.
• A gated self-attention mechanism for B-line processing, combined with sequential 1D convolutions for local enhancement, effectively models homogeneous inter-A-scan features and refines local details.

Summary by alphaXiv.