Meta learning with dual optimization for ischemic stroke lesion segmentation in brain MRI

Ischemic stroke is a serious neurological condition caused by reduced blood flow to the brain, leading to oxygen and nutrient deprivation, neuronal cell death and long-term disability. In recent years, Artificial Intelligence (AI)-driven segmentation models have been increasingly employed to assist in the automated identification of stroke lesions from brain MRI scans. However, existing segmentation models face challenges due to limited annotated datasets and difficulty in generalizing across diverse clinical scenarios. To overcome these limitations of few/ imperfect data, Meta-learning techniques have been explored by enabling the models to learn from sparsely labelled instances. Nevertheless, Meta-learning methods encounter issues such as random task sampling and suboptimal segmentation generation that jeopardize both the quality and efficiency of the segmentation process. To address these limitations, we propose a novel dual optimized Meta-learning framework i.e. ‘ OPTIM2MAML ’ that integrates a pre-optimization module for adaptive task sampling, leveraging spectral clustering with cosine similarity and a post-optimization module that refines the segmentation process with firefly optimization. The proposed OPTIM2MAML framework is analyzed on Brain MRI (ADC and DWI images) of ISLES 2017 and ISLES 2022 datasets to improve the ischemic stroke lesion segmentation. The results confirm the effectiveness of the proposed approach, highlighting its potential as a robust clinical decision-support tool in neuroimaging.