XtraLight-MedMamba for Classification of Neoplastic Tubular Adenomas

1. Problem Statement

Colorectal cancer (CRC) is a leading cause of cancer-related mortality. While screening colonoscopies detect precancerous adenomatous polyps, risk stratification remains challenging. Specifically, low-grade tubular adenomas (TAs) are typically considered low-risk; however, a subset of these lesions progresses to high-grade dysplasia or invasive carcinoma. Current histopathological assessment relies on subjective visual interpretation, which often fails to detect subtle, spatially distributed morphological cues associated with future malignant progression.

Existing deep learning solutions face two main hurdles:

• Convolutional Neural Networks (CNNs): Limited by local receptive fields, struggling to model long-range spatial dependencies critical for understanding tissue architecture.
• Transformers: Capable of capturing global context via self-attention but suffer from quadratic computational complexity, making them inefficient for gigapixel Whole Slide Images (WSIs) and unsuitable for resource-constrained clinical settings.

The authors propose a solution that balances high accuracy with ultra-low computational cost to enable precise risk stratification of low-grade TAs.

2. Methodology: XtraLight-MedMamba

The proposed framework, XtraLight-MedMamba, is an ultra-lightweight deep learning architecture designed for WSI classification. It integrates four key components to achieve a favorable trade-off between performance and parameter efficiency:

A. Architecture Overview

The model follows a six-stage sequential architecture with channel dimensions $[8, 16, 24, 32, 48, 64]$.

1. Shallow Feature Extraction (ConvNeXt Blocks): The first three stages utilize ConvNeXt blocks. These replace traditional residual bottlenecks with depthwise separable convolutions, Layer Normalization, and MLP-style channel expansion. This design captures local morphological patterns while preserving the spatial inductive bias of CNNs and mimicking the efficiency of Vision Transformers.
2. Global Context Modeling (Parallel Vision Mamba - PVM): Stages four to six employ Parallel Vision Mamba layers. These are State Space Models (SSMs) that model both short- and long-range spatial dependencies with linear computational complexity.
   • Mechanism: Input features are split into four parallel branches (each with $C/4$ channels), processed independently by Mamba blocks, and then concatenated. This allows for efficient global contextualization without increasing the channel dimension or computational burden.
3. Feature Refinement (SCAB Module): A Spatial and Channel Attention Bridge (SCAB) module is integrated to refine intermediate representations. It combines spatial attention (via pooling and convolution) and channel attention (via global average pooling and fully connected layers) to emphasize informative regions and propagate features across scales.
4. Classification (FNOClassifier): The final layer uses a Fixed Non-Negative Orthogonal Classifier.
   • Mechanism: Instead of learning weights, the classifier uses a fixed, pre-initialized orthogonal matrix with non-negative constraints.
   • Benefit: This enforces structured geometric constraints (inducing "Neural Collapse"), maximizes inter-class separation, and drastically reduces the number of trainable parameters while improving generalization and training stability.

B. Dataset

The model was evaluated on a newly curated Neoplastic Tubular Adenoma (NPTA) dataset:

• Source: High-resolution H&E-stained WSIs from patients with low-grade TAs.
• Cohorts:
  • Case: Patients who subsequently developed CRC after the initial low-grade TA diagnosis.
  • Control: Patients with low-grade TAs who did not develop CRC during follow-up.
• Scale: 135,049 high-quality tiles per class (224x224 pixels), split into 70% training, 15% validation, and 15% testing.

3. Key Contributions

1. Ultra-Lightweight Architecture: Introduction of XtraLight-MedMamba, which combines ConvNeXt and Parallel Vision Mamba to achieve state-of-the-art performance with only ~32,000 parameters.
2. Novel Architectural Strategy:
   • Integration of ConvNeXt for efficient local feature extraction.
   • Use of Parallel Vision Mamba for linear-complexity global dependency modeling.
   • Implementation of the FNOClassifier to enforce orthogonality and reduce redundancy, addressing the limitations of standard fixed classifiers in inducing neural collapse.
3. Clinical Relevance: The model specifically targets the difficult task of distinguishing between low-risk and high-risk progression within low-grade adenomas, a gap often missed by conventional histopathology.
4. Comprehensive Evaluation: Extensive ablation studies on momentum, learning rates, optimizers, and classifier configurations, alongside visual interpretability analysis.

4. Experimental Results

The model was compared against Transformer-based (ViT, Swin Transformer) and other Mamba-based variants.

• Performance Metrics:
  • Accuracy: 97.18%
  • F1-Score: 0.9767
  • Precision: 0.9666
  • Recall: 0.9717
• Parameter Efficiency:
  • XtraLight-MedMamba uses 32,073 parameters.
  • This is significantly lower than the Swin Transformer (598k parameters) and the Vision Transformer (7.4M parameters).
  • It outperformed all other tested configurations, including Mamba models with linear classifiers (which reached ~96.50% accuracy with ~53k parameters).
• Interpretability (Grad-CAM):
  • Visualizations confirmed the model focuses on biologically relevant features such as nuclear architecture, pseudostratification, and epithelial atypia, rather than background artifacts.
  • The model successfully identified subtle cytologic irregularities that distinguish cases (future CRC) from controls.

5. Significance and Conclusion

The paper demonstrates that State Space Models (SSMs), specifically Mamba, are highly effective for digital pathology when combined with lightweight convolutional backbones and structured classifiers.

• Clinical Impact: XtraLight-MedMamba offers a computationally efficient tool that can be deployed in resource-constrained environments (e.g., edge devices or clinics with limited GPU access) to assist pathologists in identifying high-risk low-grade adenomas.
• Scientific Contribution: By achieving high accuracy with minimal parameters, the work challenges the notion that massive model complexity is required for complex histopathological tasks. It highlights the potential of Neural Collapse and fixed orthogonal classifiers to improve generalization in medical imaging.
• Future Directions: The authors note that while performance is high, some misclassifications occur due to the patchy nature of dysplasia, suggesting a need for finer-grained annotations and potentially multi-scale training strategies in future iterations.

In summary, XtraLight-MedMamba represents a significant step forward in computational pathology, providing a scalable, accurate, and interpretable solution for the early risk stratification of colorectal cancer precursors.