NAS-StrokeNet: Automated Neural Architecture Search with Adaptive Scaling for Hemorrhagic and Ischemic Stroke Classification on Multi-contrast MRI

Abstract

Precise and timely discrimination between ischemic and hemorrhagic stroke is essential for effective treatment planning in acute stroke management. Although deep learning methods have shown encouraging outcomes in automated stroke identification from Magnetic Resonance Imaging (MRI), constructing ideal neural network models for stroke subtype classification is challenging because of the intricate, multi-contrast nature of MRI data and the subtle imaging features of various stroke pathologies. In this paper, NAS-StrokeNet is presented as a new framework that combines the use of neural architecture search (NAS) along with adaptive scaling methods to develop and optimize deep learning models to classify hemorrhagic and ischemic stroke from multi-contrast MRI automatically.

Sheltered in a specialized scaling controller that dynamically adjusts the network size (width, depth, resolution) based on task-oriented performance metrics and computation limitations and a differentiable architecture search method, NAS-StrokeNet is adaptable. The approach leverages domain experience through stroke-specific search space and optimisation objectives designed to prioritize clinically meaningful features while maintaining efficiency. Our adaptive scaling approach methodically investigates the trade-off between model capacity and computing needs, producing a family of models fit for use in several clinical environments with different resource limits.

Extensive validation on a sizable multi-center dataset of 5,324 patients from 12 hospitals shows that NAS-StrokeNet greatly beats both past automated methods and manually crafted topologies. With 96.8% sensitivity and 97.5% specificity, our largest model achieves 97.2% accuracy in discriminating between hemorrhagic and ischaemic stroke; our smallest compact model retains 94.3% accuracy with just 15% of the processing needs. Crucially, where traditional methods generally fail, NAS-StrokeNet shows remarkable performance in difficult circumstances like tiny lesions, early-stage strokes, and unusual presentations. The capacity of the framework to create tailored designs for particular clinical deployment situations meets a major demand in putting AI-based stroke diagnosis tools into regular clinical practice.