Edge-Enhanced MRI to CT Generation Based on Res2Net and Receptive Field Expansion

The combination of magnetic resonance imaging (MRI) and computed tomography (CT) technology integrates the high soft tissue contrast of MRI with the precise electron density information provided by CT, which is crucial for accurate dose calculations in only-MRI radiotherapy. Despite the rapid advancement of deep learning, AI-based medical image synthesis still faces challenges such as artifacts, blurred edges, and loss of detail. To address these issues, the paper proposes an MRI-to-CT translation framework that employs an edge-enhanced Res2Net for MRI-to-CT synthesis based on the principle of receptive field expansion. The algorithm uses high-frequency receptive field (HFRF) to enhance global feature extraction, introduces a multi-scale attention feature fusion method for detail enhancement, and combines jump connections and deep supervised edge-guided modules(EGM) to optimize the feature fusion strategy of the decoder. This paper conducts ablation experiments and comparative experiments between different image generation algorithms on 180 pairs of brain and pelvic datasets of the Synth-RAD 2023 competition. It performs well in quantitative metrics, with synthesized CT images more similar to actual CT images globally and locally. This research advances the synthesis of high-quality medical images from MRI to CT, providing a valuable tool for clinical radiotherapy, offering significant academic and clinical value.