[Paper Review] StrokeNeXt: A Siamese-encoder Approach for Brain Stroke Classification in Computed Tomography Imagery

Leo Thomas Ramos, Ángel D. Sappa|arXiv (Cornell University)|Feb 16, 2026

Acute Ischemic Stroke Management0 citations

TL;DR

StrokeNeXt uses a dual-branch ConvNeXt siamese encoder with a lightweight fusion decoder to classify brain stroke and its subtypes in 2D CT images, achieving state-of-the-art accuracy with efficient inference.

ABSTRACT

We present StrokeNeXt, a model for stroke classification in 2D Computed Tomography (CT) images. StrokeNeXt employs a dual-branch design with two ConvNeXt encoders, whose features are fused through a lightweight convolutional decoder based on stacked 1D operations, including a bottleneck projection and transformation layers, and a compact classification head. The model is evaluated on a curated dataset of 6,774 CT images, addressing both stroke detection and subtype classification between ischemic and hemorrhage cases. StrokeNeXt consistently outperforms convolutional and Transformer-based baselines, reaching accuracies and F1-scores of up to 0.988. Paired statistical tests confirm that the performance gains are statistically significant, while class-wise sensitivity and specificity demonstrate robust behavior across diagnostic categories. Calibration analysis shows reduced prediction error compared to competing methods, and confusion matrix results indicate low misclassification rates. In addition, the model exhibits low inference time and fast convergence.

Motivation & Objective

• Address the need for accurate and efficient CT-based stroke classification in emergency settings.
• Propose a dual-branch Siamese-encoder architecture to enhance feature diversity for stroke detection and subtyping.
• Evaluate StrokeNeXt on a real-world CT dataset and compare with CNN and Transformer baselines.
• Analyze calibration, reliability, and computational efficiency to assess clinical deployability.
• Provide a tunable model family balancing accuracy and resource usage.

Proposed method

• Dual-branch feature extraction using identical ConvNeXt encoders processing the same CT input.
• Fusion decoder that stacks encoder outputs into a synthetic sequence and merges them with a 1D convolutional fusion path and a bottleneck projection.
• A lightweight classification head operating on the fused representation.
• Training with mixed-precision and AdamW, using CrossEntropy with label smoothing, on 224x224 CT slices.
• Evaluation with metrics including accuracy, F1, AUROC, AUPRC, MCC, Brier score, ECE, and efficiency metrics (latency, throughput, memory, FLOPs).

Experimental results

Research questions

• RQ1 Can a dual-branch ConvNeXt-based Siamese encoder improve stroke detection and subtype classification over single-branch models?
• RQ2 What is the trade-off between model size, accuracy, calibration, and inference efficiency for CT-based stroke classification?
• RQ3 Does the proposed fusion decoder provide advantages over simple concatenation or summation of branch features?
• RQ4 How does StrokeNeXt perform relative to CNN and Transformer baselines on stroke presence and stroke subtype tasks?

Key findings

• StrokeNeXt variants achieve high stroke presence accuracy and F1 > 0.98 with AUROC and AUPRC around 0.99 or higher.
• StrokeNeXt-tiny and StrokeNeXt-large show strong subtype classification performance with AUROC/AUPRC ≈ 1.0 and MCC ≈ 0.973–0.986.
• Compared with baselines, StrokeNeXt-tiny achieves higher accuracy and calibration with substantially fewer parameters and lower FLOPs; StrokeNeXt-large achieves the best accuracy with higher cost.
• McNemar tests indicate statistically significant improvements over all listed baselines (p < 0.05).
• The dual-branch design yields robust per-class sensitivity/specificity, with near-perfect separation for stroke presence and high reliability for subtypes.
• StrokeNeXt offers a tunable trade-off between real-time deployment (tiny) and peak performance (large) without sacrificing calibration.