Robust Medical Image Reconstruction Using a Self-Evolving Encoder–Decoder and Adaptive Convolutional Power Scaling
Year: 2026
Type: article
Abstract: Robust medical image reconstruction is a critical requirement for accurate diagnosis and clinical decision-making, particularly when images are affected by degradation, noise, or low resolution. Conventional encoder–decoder-based reconstruction methods compress input images into low-dimensional representations and subsequently decode them into high-resolution outputs; however, such approaches often suffer from artifacts and loss of fine anatomical details under severe degradation. To address these limitations, this work proposes a robust medical image reconstruction framework using a self-evolving encoder–decoder and adaptive convolutional power scaling. The proposed super-resolution model incorporates a dynamic encoder and decoder that adaptively evolve during training to capture color contrast, structural similarity, and high-frequency details from medical images. An MLP enhanced with an adaptive power flex layer is embedded within the reconstruction pipeline, enabling learnable power-based feature scaling through weight-wise modulation and initialization. This mechanism improves feature discrimination and stabilizes the reconstruction of subtle anatomical structures. The DRIVE and CHASE_DB1 retinal image datasets are employed for experimental validation, with appropriate preprocessing applied before training and testing. The selected images are processed through the proposed super-resolution model, and performance is quantitatively evaluated using PSNR, SSIM, sensitivity, and specificity metrics. Experimental results demonstrate that the proposed method achieves significant improvements in reconstruction quality and robustness compared to existing approaches, yielding enhanced perceptual quality and structural fidelity in reconstructed medical images. These findings indicate that the proposed self-evolving encoder–decoder with adaptive convolutional power scaling is well-suited for reliable medical image reconstruction applications.
Authors: Dhanusha P B, J. Bennilo Fernandes, A. Muthukumar, A. Lakshmi
Language: English
Subfield: Computer Vision and Pattern Recognition
Field: Computer Science
Domain: Physical Sciences
Open Access status: diamond
OpenAlex ID: W7133787848
FWCI:
Cites:
Cited by:
Related to: