Saliency-Guided Implicit Neural Representation Learning for Stroke Diagnosis in Diffusion-Weighted MRI

Abstract

Early diagnosis of acute ischemic stroke from diffusion-weighted magnetic resonance imaging (DWI-MRI) remains challenging because subtle lesions often exhibit weak contrast and irregular boundaries that are difficult to capture using conventional 3D convolutional neural networks. This paper proposes a Saliency-Guided Implicit Neural Representation (INR) framework that enables continuous sub-voxel feature learning for improved stroke diagnosis. A context-conditioned INR module models features at arbitrary spatial coordinates, while a lightweight saliency prediction network adaptively allocates computational resources to diagnostically relevant regions. An attention-based feature aggregation mechanism integrates informative sub-voxel representations for robust volumetric classification. Extensive experiments on a multi-center DWI-MRI dataset demonstrate that the proposed framework outperforms conventional 3D ResNet, attention-based, and anti-aliasing baselines, achieving an AUC-ROC of 0.947 and sensitivity of 0.918 with modest computational overhead. The proposed approach preserves fine lesion textures and boundary continuity, providing a reliable and efficient solution for early computer-aided stroke diagnosis.