A Novel Multi-branch ConvNeXt Architecture for Identifying Subtle Pathological Features in CT Scans

I. Perera and U. Thayasivam

arXiv preprint arXiv:2510.09107, October 2025

---

Abstract

This repository contains the implementation and experimental results of a novel multi-branch ConvNeXt architecture designed to identify subtle pathological features in CT scans. The proposed method integrates three complementary pooling strategies—Global Average Pooling, Global Max Pooling, and Attention-Weighted Pooling—to capture both holistic contextual patterns and fine-grained diagnostic markers that conventional convolutional neural networks frequently overlook.

Key Contributions

• Multi-branch Feature Extraction: A novel architecture combining three parallel pooling pathways to extract complementary feature representations from CT scans
• Attention-Weighted Pooling Mechanism: An adaptive attention module that learns to emphasize diagnostically relevant regions, providing implicit explainability
• Comprehensive Preprocessing Pipeline: Domain-specific enhancements including CLAHE contrast adjustment, two-lung ROI extraction, and systematic data augmentation
• Two-Phase Transfer Learning Strategy: Structured training methodology with controlled fine-tuning to prevent overfitting while maintaining generalizability

Performance

The proposed architecture was evaluated on combined COVID-19 CT datasets and demonstrates state-of-the-art performance:

Metric	Value
ROC-AUC	0.9937
Accuracy	97.57%
Precision	98.35%
Recall	98.15%
F1-Score	98.25%

These results surpass existing methods including MA-Net, ResNet+CBAM, MTL, and other contemporary architectures.

Repository Structure

├── docs/               # Research documentation (proposal, literature review, methodology)
├── experiments/        # Experimental notebooks (baseline and transfer learning studies)
├── src/                # Source code and model implementation
├── results/            # Experimental results and performance visualizations
├── data/               # Dataset information and preprocessing utilities
└── assets/             # Figures, diagrams, and supplementary materials

Detailed documentation is provided within each directory for comprehensive understanding of the methodology and implementation.

Citation

If you use this work in your research, please cite:

@misc{perera2025novel,
      title={A Novel Multi-branch ConvNeXt Architecture for Identifying Subtle Pathological Features in CT Scans}, 
      author={I. Perera and U. Thayasivam},
      year={2025},
      eprint={2510.09107},
      archivePrefix={arXiv},
      primaryClass={cs.CV},
      doi={10.48550/arXiv.2510.09107}
}

Contact

For questions, issues, or collaboration inquiries, please open an issue in this repository.

---

This research was conducted as part of the Healthcare AI: Medical Imaging research program (Student ID: 210471F).