A Hybrid PNet and U-Net Autoencoder Framework for High-Fidelity Medical Image Super-Resolution

DOI: https://doie.org/10.10399/JBSE.2025923859

Nalavadi Srikantha, Chandrashekhar K

Keywords:

Medical Image Super-Resolution, U-Net, Pyramid Networks (PNet), Autoencoder, Deep Learning, Multi-scale Feature Extraction.

Abstract:

High-resolution (HR) medical imaging is critical for accurate clinical diagnosis and computer-aided analysis. However, acquiring HR images is often limited by hardware constraints, acquisition time, and patient safety concerns regarding radiation exposure. Single Image Super-Resolution (SISR) has emerged as a post-processing solution, yet traditional convolutional neural networks (CNNs) often struggle to preserve high-frequency diagnostic details, resulting in over-smoothed textures [4]. This paper proposes a novel deep learning framework that integrates a Pyramid Network (PNet) into U-Net based autoencoder architecture to address these limitations. The proposed model leverages the symmetric encoder-decoder structure of the U-Net for efficient feature reconstruction while embedding a PNet module at the bottleneck to capture multi-scale contextual information. This hybrid approach allows the network to dynamically aggregate features from different receptive fields, effectively recovering both global anatomical structures and local tissue textures. Furthermore, a residual learning [8] strategy is employed to accelerate convergence and allow the network to focus on learning high-frequency residuals rather than the direct image mapping. Extensive experiments were conducted on the BraTS 2020 and DeepLesion dataset. Quantitative evaluations demonstrate that the proposed Hybrid PNet-UNet outperforms state-of-the-art methods, achieving a Peak Signal-to-Noise Ratio (PSNR) of 31.45 dB and a Structural Similarity Index (SSIM) of 0.9246. Visual results confirm that the model significantly reduces blurring artifacts and enhances edge definition, making it a promising tool for improving diagnostic precision in medical imaging.

PDF