An interpretable deep learning method for medical image deblurring and restoration

Unsharpness or blurring in medical images remains a critical artefact that can adversely affect clinical interpretation. Medical images acquired using imaging modalities such as computed tomography (CT) and X-ray are often degraded by acquisition noise and patient motion, leading to reduced diagnostic accuracy. Recently, deep learning (DL)-based models have been widely used for deblurring medical images. However, the performance of these DL-based deblurring models depends on the choice of activation function (AF). Most of the existing DL-based deblurring models employ static AFs, such as Rectified Linear Unit (ReLU), which process all image regions uniformly and may consequently amplify noise in the background or introduce ringing artefacts. To address this, we propose Spatially Adaptive Gated Activation (SAGA), a novel dynamic operator that computes activations based on local context. SAGA generates a local sharpening signal that is adaptively modulated by a learnable, per-pixel gating mechanism. The context-aware design of SAGA enables the DL models to effectively distinguish actual anatomical boundaries from acquisition noise, thereby suppressing high-frequency artefacts while preserving clinically relevant structures. We validate the performance of SAGA against six benchmark AFs – Sigmoid, hyperbolic tangent (Tanh), rectified linear unit (ReLU), Swish, exponential linear unit (ELU), and funnel ReLU (FReLU) – using four widely used DL-based image deblurring architectures, namely Visual Geometry Group Network (VGGNet), Residual Network (ResNet), U-Net, and Enhanced Deep Super-Resolution Network (EDSR), on CT and X-ray datasets. The experimental results show that SAGA consistently outperforms conventional AFs, yielding average improvements of +6.3 dB in peak signal-to-noise ratio (PSNR) and +0.09 in structural similarity index measure (SSIM) over ReLU across all four DL-based image deblurring models. These results demonstrate that SAGA can be used as an alternative AF, which helps DL-based deblurring models produce sharper and more diagnostically reliable images. Furthermore, SAGA can be seamlessly integrated into existing DL-based deblurring frameworks without any changes to the DL architecture. • Introduce a spatially adaptive activation function to improve medical image clarity. • Use context-aware learning to separate anatomical structures from imaging noise. • Improve diagnostic image quality through adaptive deep learning analytics. • Achieve higher signal quality across multiple medical imaging models. • Support safer clinical interpretation through reliable image restoration.