SIRU-Net : Smoothing Inception Residual U-Net for Deformable Medical Image Registration

Volumetric medical image registration is critical for computer-aided diagnosis and has seen advancements with Transformer-based learning techniques. These Transformers, despite their global modeling strength, require extensive parameters and high computational power, limiting their deployment in low-resource settings. Additionally, their reliance on large datasets for effective outcomes poses a challenge. In contrast, CNN-based methods offer detailed local information but struggle with global modeling and long-distance interactions. The goal of this work is to introduce a method that provides an extended receptive field, requires fewer parameters, and delivers significant results with a limited training dataset. To achieve this, we propose the SIRU-Net (Smoothing Inception Residual U-Net) architecture, which integrates an inception module to extend the receptive field, a residual module to improve feature quality, and Gaussian convolution to improve information quality in deep stages of the network. This method has been evaluated on both inter-patient and atlas-patient datasets with limited data, demonstrating that it outperforms transformer-based methods while using only 5 % of their parameters.