StrokeNeXt: A Siamese-encoder Approach for Brain Stroke Classification in Computed Tomography Imagery

StrokeNeXt is a Siamese-encoder model utilizing dual ConvNeXt branches and a lightweight decoder that achieves state-of-the-art accuracy and robust calibration for classifying ischemic and hemorrhagic strokes in 2D CT images, significantly outperforming existing convolutional and Transformer-based baselines.

The Gist

Imagine your brain is a bustling city, and blood vessels are the highways delivering oxygen and fuel. A stroke happens when a highway is either blocked by a traffic jam (an ischemic stroke) or when a road suddenly bursts, flooding the area (a hemorrhagic stroke).

Doctors use CT scans (like high-tech X-ray photos) to see what's wrong. But reading these photos is hard, slow, and requires highly trained experts. If a doctor misses a detail or confuses a "blocked road" with a "burst road," the treatment could be wrong, leading to severe damage.

This paper introduces StrokeNeXt, a new AI "detective" designed to look at these brain photos and instantly tell doctors:

1. Is there a stroke? (Yes/No)
2. What kind is it? (Blocked road vs. Burst road)

Here is how it works, explained with simple analogies:

1. The "Double-Check" Team (The Siamese Encoder)

Most AI models look at a picture with just one pair of "eyes." If that one pair misses a tiny clue, the AI gets it wrong.

StrokeNeXt is different. It uses two identical AI brains (called ConvNeXt encoders) working side-by-side on the same image.

• The Analogy: Imagine two expert detectives looking at the same crime scene photo. They aren't sharing notes; they are working independently. One might notice a faint shadow, while the other spots a tiny crack. Because they are independent, they catch different clues.
• The Result: Instead of relying on a single opinion, the system combines their unique observations to get a much clearer picture. This reduces the chance of missing something important.

2. The "Lightweight Translator" (The Fusion Decoder)

Once the two detectives have gathered their clues, they need to combine them into a final report.

• The Problem: Usually, combining two big reports makes a huge, heavy document that takes forever to read.
• StrokeNeXt's Solution: It uses a "lightweight decoder." Think of this as a super-efficient translator who takes the two detectives' notes, merges the most important parts, and summarizes them into a tiny, crisp "Yes/No" or "Type A/Type B" answer. It does this without getting bogged down in unnecessary details, making it incredibly fast.

3. The Race Against Time (Speed and Accuracy)

In a real hospital, time is life. A model that is 99% accurate but takes 10 seconds to think is useless if the doctor needs an answer in 1 second.

• The Performance: The authors tested StrokeNeXt on nearly 7,000 brain scans.
  • Accuracy: It got the diagnosis right 98.8% of the time.
  • Speed: It can analyze a brain scan in just 2 milliseconds (that's faster than a human blink).
  • Comparison: It beat other famous AI models (like ResNet, VGG, and even complex "Transformer" models) that were either slower, less accurate, or both.

4. Why This Matters

• No More Guessing: The model is so good at distinguishing between a "blocked road" and a "burst road" that it rarely makes mistakes. This is crucial because the medicine for a blockage is the opposite of the medicine for a burst. Giving the wrong medicine can be fatal.
• Accessible: Because the model is so efficient, it doesn't need a supercomputer to run. It could potentially run on standard hospital equipment, helping doctors in smaller clinics or busy emergency rooms make faster, better decisions.

The Bottom Line

StrokeNeXt is like hiring a team of two super-observant detectives who work independently, then quickly summarize their findings for a doctor. It is faster, smarter, and more reliable than previous AI tools, offering a powerful new tool to save lives by catching strokes early and correctly.

Technical Summary

1. Problem Statement

Brain stroke is a leading cause of mortality and disability, requiring rapid and accurate diagnosis to determine appropriate treatment (e.g., distinguishing between ischemic and hemorrhagic strokes, which have opposing treatment protocols).

• Challenges: Manual interpretation of CT scans is time-sensitive, labor-intensive, and prone to human error and inter-observer variability.
• Existing AI Limitations:
  • Traditional ML: Relies on handcrafted features and lacks spatial context.
  • Standard CNNs: Often limited in representational capacity and robustness.
  • Transformers (e.g., ViT, Swin): While powerful, they typically require massive datasets, high computational costs, and long training times, making them less suitable for resource-constrained or time-critical clinical environments.
  • 3D Approaches: Exploit volumetric data but are computationally expensive and prone to overfitting.
• Goal: Develop a model that balances high diagnostic accuracy, robust calibration, and computational efficiency for real-time clinical deployment.

2. Methodology: StrokeNeXt Architecture

The proposed StrokeNeXt is a Deep Learning model designed for 2D CT image classification. It utilizes a dual-branch Siamese encoder architecture coupled with a lightweight decoder.

A. Dual-Branch Siamese Encoder

• Structure: Two identical ConvNeXt encoders process the same input image in parallel.
• Hypothesis: Although the encoders share the same input, they do not share weights. Independent stochastic training trajectories allow each branch to specialize in different visual cues and abstraction paths, capturing complementary representations and mitigating information loss common in single-path pipelines.
• Backbone: Based on ConvNeXt, a CNN architecture inspired by Transformer design principles (using large 7×7 kernels, GELU activations, and layer normalization) but retaining the computational efficiency of CNNs. The authors utilize variants ranging from tiny to large.

B. Lightweight Fusion Decoder

Instead of simple concatenation or element-wise summation, the model employs a custom fusion module:

1. Stacking: Outputs from the two encoders are stacked to form a synthetic sequence of length 2 ($[B, C, 2]$).
2. 1D Convolution: A 1D convolution with kernel size $k=2$ collapses the sequence dimension. This learns local interactions and cross-channel mixing between the two branches, effectively capturing cross-branch dependencies without the overhead of attention mechanisms.
3. Bottleneck Transformation: The fused features pass through a bottleneck layer (two pointwise convolutions with an optional hidden dimension $H$) followed by BatchNorm, GELU, and Dropout. This refines the representation and controls model capacity without modifying the encoders.
4. Classification Head: A compact head generates the final class probabilities.

C. Dataset and Training

• Dataset: 6,774 annotated brain CT images (PNG format) from the TEKNOFEST 2021 competition, curated by seven radiologists.
  • Classes: Non-stroke (4,551), Hemorrhagic (1,093), Ischemic (1,130).
• Tasks:
  1. Stroke Detection: Binary classification (Stroke vs. Non-stroke).
  2. Subtype Classification: Multi-class classification (Ischemic vs. Hemorrhagic).
• Training: Implemented in PyTorch with Mixed Precision (AMP). Optimized using AdamW with label smoothing. Data augmentation included flips, rotations, and jittering.

3. Key Contributions

1. Novel Architecture: Introduction of StrokeNeXt, a dual-branch ConvNeXt framework that enhances feature diversity through parallel processing while maintaining efficiency.
2. Efficient Fusion: A custom lightweight decoder using 1D convolutions and bottleneck transformations that outperforms standard concatenation or attention-based fusion in terms of parameter count and latency.
3. State-of-the-Art Performance: Demonstrated superior accuracy and F1-scores compared to both traditional CNNs (ResNet, VGG, MobileNet) and Transformer-based models (Swin, ViT).
4. Robust Calibration: The model exhibits low Expected Calibration Error (ECE) and Brier scores, ensuring reliable probability estimates crucial for clinical decision-making.
5. Comprehensive Benchmarking: Established a new baseline for CT-based stroke classification, validated through statistical significance testing (McNemar's test) across multiple metrics.

4. Experimental Results

The model was evaluated on two tasks with variants ranging from tiny to large.

A. Stroke Presence Detection (Stroke vs. Non-Stroke)

• Accuracy: StrokeNeXt-Large achieved 98.7% accuracy and an F1-score of 0.987.
• Comparison: Significantly outperformed baselines:
  • MobileNetV2: 86.5% accuracy.
  • Swin Transformer: 89.3% accuracy.
  • ConvNeXt-base (single branch): 87.9% accuracy.
• Efficiency: The StrokeNeXt-tiny variant achieved 97.8% accuracy with only 57.6M parameters and 8.98 GFLOPs, processing 570 images/second with a latency of 2ms.
• Statistical Significance: McNemar's tests confirmed that performance gains over all baselines were statistically significant ($p < 0.0001$).

B. Stroke Subtype Classification (Ischemic vs. Hemorrhagic)

• Accuracy: StrokeNeXt-Base achieved 98.8% accuracy and an F1-score of 0.988.
• Sensitivity/Specificity: Achieved near-perfect separation (>0.99) for both classes, avoiding the class imbalance issues seen in ResNet and VGG models (which often over-predicted non-stroke or hemorrhagic cases).
• Comparison: Outperformed previous SOTA methods like OzNet (0.984 F1) and D-UNet (0.985 F1).
• Clinical Relevance: The model correctly identified over 100 ischemic cases where other models failed, minimizing false negatives in critical diagnostic scenarios.

5. Significance and Conclusion

• Clinical Impact: StrokeNeXt addresses the critical need for rapid, accurate, and reliable stroke diagnosis. Its ability to distinguish between ischemic and hemorrhagic strokes with high sensitivity reduces the risk of inappropriate treatment (e.g., administering thrombolytics to a hemorrhagic patient).
• Deployment Viability: Unlike heavy Transformer models, StrokeNeXt offers a tunable trade-off between performance and efficiency. The tiny variant is suitable for real-time, resource-limited environments, while the large variant maximizes accuracy for high-end systems.
• Generalizability: The architecture generalizes effectively across different diagnostic tasks (detection vs. subtyping) without requiring 3D volumetric data, making it practical for standard 2D CT workflows.

In summary, StrokeNeXt represents a significant advancement in medical AI by successfully integrating the representational power of dual-branch learning with the computational efficiency required for real-world clinical deployment.