Joint Segmentation-Registration Optimization for Single Snapshot Spatial Frequency Domain Imaging

Abstract

Multiwavelength spatial frequency domain imaging (SFDI) technology faces challenges in clinical applications, including background redundancy and motion artifacts, which compromise the stability of physiological parameter extraction. To address this, this article proposes a segmentation-registration joint optimization method based on single snapshot SFDI: during the preprocessing stage, a segmentation model is employed to remove redundant background; During registration, AR-Net performs deformable registration and image resampling for motion-shifted tissue images, thereby reducing measurement biases induced by motion and geometric variations, such as depth rotation ( $\le$ $ 15^{\circ }$ ). In experiments using a polytetrafluoroethylene (PTFE) foot phantom, the measurement bias in apparent diffuse reflectance estimation after registration decreased to 1.76%~8.91% within 15° of horizontal and vertical rotation; static blood oxygen saturation measurement error decreased from 22.6% to 8.18%, with effective suppression of motion artifacts. Arterial occlusion experiments further validated the method’s stability and validity for measuring dynamic physiological parameters. This study provides a reliable solution for the clinical application of multiwavelength SFDI technology.