Residual-corrected deep learning framework for cerebral blood flow estimation in diffuse correlation spectroscopy

Diffuse correlation spectroscopy (DCS) is a critical non-invasive technique for cerebral blood flow monitoring, however its accuracy is frequently impaired by extracerebral or superficial layer interference. In this study, we propose a residual-corrected deep-learning framework specifically designed to stabilize cerebral blood flow estimation. By incorporating a residual learning architecture, the model effectively captures the subtle deviations between theoretical analytical solutions and experimental measurements, thereby improving the robustness of feature extraction from autocorrelation signals. We validated the proposed framework using a two-layer blood flow phantom dataset across three source-detector separations (SDS) of 1, 2 and 3 cm. Ablation experiments were further performed to verify the effectiveness of the key modules, including dual-branch architecture, attention mechanism, residual correction module, and three-stage training strategy. The experimental results demonstrate that the proposed method significantly reduces estimation errors and enhances the reliability of deep-layer blood flow quantification.