AdaSurvMamba

Dynamic Fusion and Semantic Scanning for Multimodal Survival Analysis

Overview

AdaSurvMamba combines WSI patch features with six genomic signature groups for multimodal survival analysis.

• Dual-Scale Importance-Aware Reconstruction (DSIR) adapts cross-modal interactions at both patient and token levels.
• Semantic Aggregation Scanning (SAS) organizes related tokens through learned prototypes and modulates the Mamba state-transition step size.
• Reproducible release includes BRCA/UCEC genomic profiles, five-fold splits, WSI feature downloads, and ready-to-run training scripts.

Quick Start

1. Create the environment

git clone https://github.com/zjlGO/AdaSurvMamba.git
cd AdaSurvMamba
conda env create -f environment.yml
conda activate AdaSurvMamba

The environment uses Python 3.10, PyTorch 2.0.1, and CUDA 11.8. If causal-conv1d or mamba-ssm must be built from source, CUDA 11.8 with nvcc is required.

Installation note: Mamba dependencies are sensitive to CUDA, compiler, and PyTorch versions. If setup fails, we recommend using Codex to diagnose and install the environment.

2. Download the WSI features

The TCGA-BRCA and TCGA-UCEC genomic profiles are already included in dataset/gene/. Download the matching UNI WSI features:

Cohort	WSI features	Access code
TCGA-BRCA	Baidu Netdisk	a1e3
TCGA-UCEC	Baidu Netdisk	wi7w

Extract both downloaded archives without renaming their cohort directories:

unzip /path/to/tcga_brca.zip -d dataset/WSI
unzip /path/to/tcga_ucec.zip -d dataset/WSI

Keep the genomic ZIP files compressed. The completed data directory must be:

dataset/
├── gene/
│   ├── signatures.csv
│   ├── tcga_brca_all_clean.csv.zip
│   └── tcga_ucec_all_clean.csv.zip
└── WSI/
    ├── tcga_brca/uni_20x/pt_files/*.pt
    └── tcga_ucec/uni_20x/pt_files/*.pt

3. Train

After preparing the data, run both cohorts and all five folds directly:

bash train.sh

Select cohorts or a GPU when needed:

bash train.sh brca
bash train.sh ucec
CUDA_VISIBLE_DEVICES=1 bash train.sh brca ucec

An external dataset root with the same gene/ and WSI/ layout is also supported:

bash train.sh --data-root /path/to/dataset brca

For a single fold, use the Python entry point:

python main.py --cohort brca --folds 0

Repository Layout

AdaSurvMamba/
├── data_loading/       # Survival data loading code
├── dataset/            # Genomic data and downloaded WSI features
├── mamba_ssm/          # Semantic Mamba implementation
├── models/             # AdaSurvMamba architecture
├── splits/5foldcv/     # BRCA and UCEC cross-validation splits
├── utils/              # Training and evaluation utilities
├── environment.yml
├── main.py
└── train.sh

Citation

If this work is useful in your research, please cite:

@inproceedings{zhong2026adasurvmamba,
  title={AdaSurvMamba: Dynamic Fusion and Semantic Scanning for Multimodal Survival Analysis},
  author={Zhong, Jialong and Liu, Tingwei and Yue, Baokun and Li, Jingjing and Piao, Yongri and Zhang, Miao and Liu, Leiye and Jiang, Jiahong and Ji, Wei and Lu, Huchuan},
  booktitle={Medical Image Computing and Computer Assisted Intervention -- MICCAI},
  year={2026}
}

Acknowledgements

The survival preprocessing and evaluation protocol builds on MCAT. The modality-wise mean aggregation and bottleneck survival head follow SurvPath. The selective scan implementation is adapted from Mamba.

License

This project is released under the GPL-3.0 License.

If you have any queries, please contact us at jialongzhong@mail.dlut.edu.cn.