Stroke type classification struggles with real-world data shifts and noise
Acute Stroke Classification Using Asymmetric Data Standardization and Orthogonal-Focal Loss Optimized CNN-ViT Cross-Attention Fusion: A Multicenter External Validation Study
Artificial Intelligence
Summary
Quickly telling whether someone has a stroke caused by a clot or by bleeding is crucial for treatment. The authors created a computer system that looks at brain scans using new image processing and combined computer vision methods. Although it worked very well on data it was trained on, the system performed less well on new data from different hospitals and machines. The study shows that such models can be overly confident but miss some stroke types, highlighting the challenge of making reliable AI tools for diverse medical settings.
What this means in practice
- •For hospital imaging teams: Improve stroke subtype identification workflows by integrating cross-attention fusion models that combine CNN and ViT features for CT scans.
- •For medical device developers: Design AI-powered diagnostic tools that better handle hardware variability and domain shifts in stroke imaging data using orthogonal-focal loss optimized fusion methods.$Commercial implications: Enables robust AI software in stroke diagnosis devices sold to hospitals for accurate acute triage.
Authors
Tansel Uyar
Abstract
Background/Objectives: Rapid differentiation of ischemic and hemorrhagic stroke subtypes is vital for acute management. Although current deep learning models achieve high single-dataset accuracy, they suffer from "shortcut learning" and domain shift in external validation. This study proposes a hierarchical computer-aided diagnostic (CAD) system mimicking emergency triage. Methods: We introduce a 3-channel Radiology-Focused Preprocessing pipeline that feeds radiological Hounsfield Unit (HU) windows into Convolutional Neural Network (CNN) and Vision Transformer (ViT) backbones via asymmetric formats (.mat and .png). The architecture utilizes a novel joint cross-attention fusion network optimized by an Integrated Orthogonal-Focal Loss. Results: Our system achieved a 95.48% F1-score on the internal stroke dataset. However, multi-center external validation (AISD, BHSD, PhysioNet) using a Mask Anchor filter revealed a drop to the 58-62% range, exposing the "internal validation illusion" caused by hardware noise. Critically, models exhibited high precision (77%) but low sensitivity (57%) for ischemic stroke, proving prone to hemorrhage overdiagnosis. Multi-seed variance analysis demonstrated that the DenseNet201+ViT backbone yielded the lowest deviation (0.59) and highest stability, making it the most reliable architecture for real-world clinical integration. Conclusions: To overcome diagnostic challenges, future research must focus on multimodal architectures that combine 3D volumetric extraction with clinical data.