Attention機構とフーリエ損失を備えた深層ラプラシアンピラミッドネットワークを用いた構造化照明顕微鏡再構成のためのハイブリッドモデル

Structured illumination microscopy (SIM) enables superresolution imaging of biological samples but suffers from artefacts, noise, and loss of high-frequency details in low-light conditions. These problems arise due to limitations in traditional reconstruction methods such as single-scale upsampling and pixel-wise losses that fail to capture SIM's multi-scale frequency patterns. We propose Att-SIM-LapSRN, a hybrid deep learning framework that integrates Attention U-Net with Laplacian pyramid super-resolution network (LapSRN) to address these challenges. Attention gates at skip connections selectively enhance salient feature representations corresponding to moiré patterns while attenuating background noise, producing sharper reconstructions precise localisation of cell structures. The LapSRN component employs progressive multiscale upsampling across pyramid levels to reduce the bicubic interpolation. Additionally, we introduce an FFT-based loss function that explicitly targets spatial frequency patterns, ensuring structural consistency, contrast enhancement and edge sharpness critical for SIM imaging. Our model was evaluated on the BioSR dataset, demonstrating superior performance over state-of-the-art methods, with significant improvements in PSNR, SSIM, and perceptual quality metrics. Att-SIM-LapSRN achieves enhanced lateral resolution and structural fidelity, making it a robust solution for high-quality SIM reconstruction in biological imaging applications.