Deep Learning Framework for Musculoskeletal Photoacoustic Image Generation and Enhancement

To address the challenges of scarce clinical data and complex multilayer tissue structures in musculoskeletal photoacoustic imaging (PAI), this study proposes a deep learning framework integrating pseudo-data generation and super-resolution reconstruction. First, we develop MR2PA-CycleGAN for cross-modality translation using unpaired Magnetic Resonance Imaging(MRI)-PAI data, incorporating perceptual loss to preserve anatomical structures during pseudo-PAI generation. Second, the lightweight PA-FSRCNN model enhances detail recovery from low-resolution images through attention mechanisms. Experimental results demonstrate that MR2PA-CycleGAN achieves PSNR 29.42 dB and SSIM 0.56 on joint and tumor datasets, while PA-FSRCNN attains 47.37 dB PSNR at ×2 upsampling with 35% fewer parameters than conventional models. This framework provides a scalable solution to overcome data limitations and improve musculoskeletal PAI quality.