Spatial-Semantic Mamba (SS-Mamba)

Official implementation of "Spatial-Semantic Mamba: Preserving 2D Structure and Meaningful Sequencing for Enhanced Computational Pathology"

Paper

📄 Spatial-Semantic Mamba: Preserving 2D Structure and Meaningful Sequencing for Enhanced Computational Pathology

• PDF: Spatial_Semantic_Mamba.pdf

This work was developed as a graduate-level research project and is currently an unpublished manuscript / preprint.

Overview

SS-Mamba is a novel framework for Whole Slide Image (WSI) classification that addresses two critical limitations in current computational pathology methods:

1. Spatial Discrepancy: Standard methods flatten 2D patches into 1D sequences, destroying spatial relationships
2. Random Patch Ordering: MIL frameworks process patches in arbitrary order, ignoring semantic relationships

Our approach introduces:

• 2D-SSM Backbone: Bidirectional scanning along horizontal and vertical axes
• Semantic Pre-ordering: Parameter-free ordering based on feature similarity
• BiMamba Aggregator: Bidirectional Mamba for MIL aggregation

Project Structure

SS-Mamba/
├── README.md
├── requirements.txt
├── setup.py
├── configs/
│   └── config.yaml              # Configuration file
├── ss_mamba/                    # Main package
│   ├── models/
│   │   ├── ss_mamba.py          # Full SS-Mamba model
│   │   ├── backbone/
│   │   │   └── mamba_2d.py      # 2D-SSM backbone
│   │   ├── ordering/
│   │   │   └── semantic_ordering.py
│   │   └── aggregator/
│   │       └── bimamba_mil.py   # BiMamba aggregator
│   ├── data/
│   │   ├── wsi_dataset.py       # WSI dataset class
│   │   ├── preprocessing.py     # Patch extraction
│   │   └── transforms.py        # Augmentations
│   └── utils/
│       ├── metrics.py           # Evaluation metrics
│       ├── logger.py            # Logging utilities
│       └── visualization.py     # Visualization tools
├── tools/
│   ├── train.py                 # Training script
│   ├── evaluate.py              # Evaluation script
│   └── extract_features.py      # Feature extraction
└── scripts/
    ├── run_experiment.sh        # Run single experiment
    └── run_all_datasets.sh      # Run all experiments

1. Data Preparation

Download datasets and organize as follows:

data/
├── camelyon16/
│   ├── train/
│   │   ├── slides/              # .tif files
│   │   └── labels.csv
│   └── test/
├── tcga_brca/
│   ├── slides/
│   └── labels.csv
└── bracs/
    ├── slides/
    └── labels.csv

2. Patch Extraction

python tools/extract_features.py \
    --data_path /path/to/slides \
    --output_path /path/to/patches \
    --patch_size 256 \
    --magnification 20

3. Training

# Camelyon16 (binary classification)
python tools/train.py \
    --dataset camelyon16 \
    --data_path /path/to/data \
    --num_classes 2 \
    --epochs 100

# TCGA-BRCA (4-class classification)
python tools/train.py \
    --dataset tcga_brca \
    --data_path /path/to/data \
    --num_classes 4 \
    --epochs 100

# BRACS (7-class classification)
python tools/train.py \
    --dataset bracs \
    --data_path /path/to/data \
    --num_classes 7 \
    --epochs 150

Or use shell scripts:

bash scripts/run_experiment.sh camelyon16
bash scripts/run_all_datasets.sh

4. Evaluation

python tools/evaluate.py \
    --dataset camelyon16 \
    --checkpoint /path/to/checkpoint.pth \
    --data_path /path/to/data

Configuration

Edit configs/config.yaml for custom settings:

model:
  hidden_dim: 512
  num_layers: 2
  dropout: 0.25

ordering:
  lambda_spatial: 0.5
  sigma: 100

training:
  lr: 0.0002
  weight_decay: 0.05
  epochs: 100
  batch_size: 1

Datasets

Dataset	WSIs	Classes	Task	Link
Camelyon16	399	2	Metastasis Detection	Link
TCGA-BRCA	1,098	4	Molecular Subtyping	Link
BRACS	547	7	Tumor Typing	Link

Acknowledgments

• Mamba for the State Space Model implementation
• VMamba for 2D scanning patterns
• CLAM for WSI preprocessing pipeline

Contact

• Sina Mansouri - Smansou3@gmu.edu
• Neelesh Prakash Wadhwani - nwadhwan@gmu.edu
• Philip Stavrev - pstavrev@gmu.edu