Enhancing Ischemic Stroke Analysis with Multi-Scale Feature Extraction and Early Fusion in a Deep Learning Framework

Ischemic stroke diagnosis and treatment planning demand accurate and efficient lesion segmentation and classification. Existing methods often rely on either handcrafted features or deep learning models, but their performance can be limited due to incomplete feature representation or insufficient training data. To address these limitations, we propose a novel framework that combines handcrafted and deep features extracted from multimodal MRI (DWI, FLAIR, T1), along with relevant clinical data. Our approach leverages a pre-trained 3D ResNet model for deep feature extraction, capturing complex patterns within the MRI data, while handcrafted features provide domain-specific insights into lesion characteristics. We utilize early fusion to integrate these diverse feature sets, employing an attention mechanism to dynamically weight their importance. The fused feature vectors are then input into a Random Forest classifier for accurate and interpretable prediction of ischemic stroke. This multi-scale approach, incorporating both traditional and deep learning techniques, offers a comprehensive and robust representation of ischemic stroke, potentially improving the accuracy and efficiency of diagnosis in clinical practice. The proposed pipeline is trained and evaluated on its own collected dataset of 500 patient cases with expert annotations serving as ground truth. Our method achieves promising results in terms of lesion segmentation accuracy (Dice Similarity Coefficient: 0.80) and classification performance (accuracy: 0.95, AUC-ROC: 0.97). Additionally, we explore the impact of different fusion strategies and the inclusion of clinical features on model performance. Our findings demonstrate the potential of this integrated approach for enhancing ischemic stroke analysis in clinical settings, potentially leading to faster and more accurate diagnosis, treatment planning, and ultimately, improved patient outcomes.