基于双注意力引导特征级联的显微影像深度估计方法

Dual-Attention Guided Feature Cascade Method for Microscopic Image Depth Estimation

深度估计在医学显微影像中具有重要应用价值,可以弥补外科医生在手术过程中由于观察目镜感官受限而难以获得精确深度信息的不足。针对手术场景动态多变、软组织和手术器械尺度微小等原因导致深度估计精度不高的问题,提出了一种改进稠密回归中跨层级特征级联的深度估计方法。通过利用多层次特征聚合模块,将编码器中的上下文信息传递到解码器中,同时基于通道选择和分支优化的双重注意力特征融合机制来优化解码的精度。为了获得密集的深度真值,提出了一种迭代式配准策略,结合自动化的机械臂扫描实现由粗到精优化多视角点云配准,并从模拟场景中重建高精度深度数据。结果表明,本文提出的深度估计方法实现了0.001 51的均方误差值(root mean squared error,RMSE)和0.030 39的尺度不变对数误差值(scale-invariant log,SILog),超越了以往最先进的方法,并对细小手术器械的尖端产生了更精准的深度估计。

Depth estimation has significant applications in medical microscopic imaging,as it can compensate for the limitations of surgeons in obtaining accurate depth information due to the sensory restrictions of the surgical microscope.To address the problem of low depth estimation accuracy caused by factors such as dynamic changes in surgical scenes,soft tissues,and small-scale surgical instruments,a novel depth estimation method is proposed to enable cross-level feature cascade in dense regression.By employing a multi-level feature aggregation module,the contextual information from the encoder was propagated to the decoder.A dual-attention guided feature fusion mechanism based on channel selection and branch optimization was designed to enhance the decoding accuracy.To obtain dense depth ground-truth values,an iterative registration strategy with an automated robotic scanning approach was introduced,which optimized the multi-view point cloud registration process from coarse to fine and reconstructed high-precision depth data from simulated scenes.Experimental results show that the proposed depth estimation method achieves a root mean squared error(RMSE) value of 0.001 51 and a scale-invariant log error(SILog) value of 0.030 39,which surpasses previous state-of-the-art methods and produces a more accurate depth distribution for the tiny surgical instrument tip.