EWT-HA-Net: An Efficient Wavelet-Convolution-Enhanced Hybrid Attention Network for Sensorless Freehand 3D Ultrasound Reconstruction

Abstract

Sensorless freehand three-dimensional (3D) ultrasound reconstruction eliminates the need for external tracking devices but remains challenging due to speckle noise, weak textures, ambiguous anatomical boundaries, and the computational cost of existing high-performance methods. To address these issues, this paper proposes EWT-HA-Net, an efficient wavelet-convolution-enhanced hybrid attention network for sensorless freehand 3D ultrasound reconstruction. The network estimates inter-frame transformations by exploiting the selected frame pair and sequence contextual information, and recovers the probe trajectory through sequential transformation accumulation. An Enhanced Wavelet Transform Convolution (EWTConv) module integrates learnable wavelet decomposition and Dynamic Frequency Fusion Gating (DFFG) for multi-scale spatial-frequency feature extraction, while a complementary CoordAtt-SE attention strategy enhances position-sensitive and channel-wise feature representation. Experimental results demonstrate that EWT-HA-Net achieves improved reconstruction performance with low computational complexity, providing a favorable balance between accuracy and efficiency.