A LIGHTWEIGHT MULTIMODAL DEEP LEARNING FRAMEWORK FOR ACCURATE SEGMENTATION OF ACUTE ISCHAEMIC STROKE LESIONS

Abstract

Accurate delineation of acute ischaemic stroke lesions remains a significant challenge in neuroimaging due to substantial variation in lesion appearance, modality-specific contrast and anatomical complexity. This study introduces a lightweight multimodal deep learning framework that integrates diffusion-weighted imaging, apparent diffusion coefficient maps and fluid-attenuated inversion recovery sequences to capture complementary aspects of acute infarction within a unified representation. The architecture combines efficient feature extraction with context-aware modelling to characterise anisotropic lesion morphology while preserving fine spatial detail during reconstruction. Evaluation on a subject-independent subset of the ISLES 2022 dataset demonstrated consistently high spatial agreement with expert annotations, achieving a Dice coefficient of 0.93, an intersection-over-union of 0.88, a precision of 0.95, a recall of 0.92 and an F1 score of 0.93. These results indicate that the framework reliably identifies diffusion-restricted tissue while maintaining low false-positive rates across diverse imaging conditions. Qualitative inspection further confirmed coherent lesion boundaries and stable localisation even in cases involving subtle, fragmented or low-contrast lesions. Overall, the findings show that an efficient multimodal architecture can achieve segmentation performance comparable to substantially larger models, offering a practical and scalable solution for automated stroke-lesion analysis.