PromptM-UNet: Efficient Text-Prompted 3D Medical Image Segmentation with Mamba SSM

PromptM-UNet is an ultra-lightweight, multi-modal 3D medical image segmentation framework that integrates Residual Vision Mamba (SSM) with multi-stage language conditioning. It enables accurate, prompt-guided volumetric CT segmentation while maintaining linear (O(N)) computational complexity and a minimal memory footprint feasible for standard consumer GPUs.

• GitHub Repository: https://github.com/kiuyha/PromptM-UNet
• Target Organ: Spleen (TotalSegmentator CT Cohort)

---

🌟 Key Features

• Linear (O(N)) Volumetric Modeling: Built on 3D Residual Vision Mamba (RVM) blocks from LightM-UNet, providing global receptive field modeling without quadratic (O(N^2)) Transformer memory explosion.
• Ultra-Lightweight Footprint: Active visual backbone and projection heads require only ~5.11M parameters, with complete end-to-end model footprint of only ~27.83M parameters when using Sentence-BERT ((16\times) smaller than standard 3D nnU-Net, (100\times) smaller than SegVol).
• Zero-Overhead Language Conditioning: Decoupled text embedding caching ensures frozen language models (Sentence-BERT, CLIP, BioBERT) consume 0 extra GPU memory during training/inference.
• Dual-Resolution Spatial Zoom: SegVol-inspired dual-scale pipeline (3mm global context + 1.5mm high-resolution target crop).
• Clinical Performance: Achieves (0.870+) Dice Similarity Coefficient (DSC) and (0.793+) Normalized Surface Distance (NSD) on TotalSegmentator CT scans with under 4.3 GB validation VRAM.

---

📊 Benchmark & Ablation Results

The following results are evaluated on the official TotalSegmentator spleen validation cohort across 30 training epochs (from notebooks/ablation_promptm_unet.ipynb):

Parameter Accounting & Model Weights: The checkpoint files (.pth) store only the trainable visual backbone, stage projection layers, squeeze convolutions, and deep supervision heads (~5.03M to ~5.44M parameters). Frozen language encoder weights are decoupled and loaded dynamically:

• CLIP (openai/clip-vit-base-patch32): 63.17M parameters (Default text encoder used for Ablations 1–3)
• BioBERT (emilyalsentzer/Bio_ClinicalBERT): 108.31M parameters
• Sentence-BERT (SBERT) (sentence-transformers/all-MiniLM-L6-v2): 22.71M parameters

The Total Parameters column indicates the full end-to-end model (Visual Backbone + Fusion Heads + Text Encoder).

1. Fusion Operation Ablation

Evaluated with Standard BCE (50:50), default CLIP text encoder (63.17M), and All-Stage Fusion:

Operation	Trainable Params	Total Params	Peak Epoch	Peak DSC	Peak NSD
Multiplication	5.26M	68.43M	24	0.8619 ± 0.0028	0.7871 ± 0.0041
Concatenation	5.44M	68.60M	26	0.8381 ± 0.0000	0.7647 ± 0.0000
Hybrid	5.44M	68.60M	28	0.8152 ± 0.0014	0.7301 ± 0.0025

2. Fusion Strategy Ablation

Evaluated with Standard BCE (50:50), Multiplication operation, and default CLIP text encoder (63.17M):

Fusion Strategy	Trainable Params	Text Encoder	Total Params	Peak Epoch	Peak DSC	Peak NSD	Peak Val VRAM
Early Fusion	5.15M	CLIP (63.17M)	68.31M	19	0.8622 ± 0.0012	0.7928 ± 0.0019	< 4.3 GB
All-Stage Fusion	5.26M	CLIP (63.17M)	68.43M	24	0.8619 ± 0.0033	0.7871 ± 0.0047	< 4.3 GB
Late Fusion	5.03M	CLIP (63.17M)	68.20M	12	0.8576 ± 0.0008	0.7675 ± 0.0013	< 4.3 GB

3. Text Encoder Ablation

Evaluated with Standard BCE (50:50), Multiplication operation, and Early Fusion:

Text Encoder	Pretrained Model	Text Dim	Text Encoder Params	Trainable Params	Total Params	Peak Epoch	Peak DSC	Peak NSD
Sentence-BERT (SBERT)	all-MiniLM-L6-v2	384D	22.71M	5.11M	27.83M	24	0.8702 ± 0.0009	0.7896 ± 0.0019 (0.7938 ± 0.0020*)
CLIP	clip-vit-base-patch32	512D	63.17M	5.15M	68.31M	19	0.8622 ± 0.0012	0.7928 ± 0.0019
BioBERT	Bio_ClinicalBERT	768D	108.31M	5.21M	113.52M	19	0.8466 ± 0.0016	0.7584 ± 0.0024

*Note: SBERT attained peak DSC of 0.8702 ± 0.0009 at Epoch 24, and peak NSD of 0.7938 ± 0.0020 at Epoch 26.

Multi-Tier Clinical Prompt Robustness

Evaluated on Best Early Fusion + SBERT checkpoint across prompt complexity tiers:

Prompt Tier	Clinical Description	Peak DSC
Tier (N)	Organ Name (e.g., "spleen")	0.8702
Tier (NS)	Name + Synonym (e.g., "spleen, lien")	0.8702
Tier (NL)	Name + Location (e.g., "spleen in left upper quadrant")	0.8702
Tier (NSL)	Name + Synonym + Location (e.g., "spleen, lien in upper left abdomen")	0.8703

---

🚀 Quickstart & Inference

1. Installation

# Clone the repository
git clone https://github.com/kiuyha/PromptM-UNet.git
cd PromptM-UNet

# Install dependencies and package
pip install -e .

2. Python Inference

import torch
import yaml
from promptm_unet.models.PromptMUNet import PromptMUNet

# Load configuration
with open("configs/default.yml", "r") as f:
    config = yaml.safe_load(f)

# Instantiate model
device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
model = PromptMUNet(config).to(device)

# Load checkpoint
checkpoint = torch.load("best_model.pth", map_location=device)
model.load_state_dict(checkpoint["model_state_dict"])
model.eval()

# Dummy 3D input: (Batch, Channels, Depth, Height, Width)
dummy_ct = torch.randn(1, 1, 64, 128, 128).to(device)
prompt = ["spleen segmentation in abdominal CT scan"]

# Forward pass
with torch.no_grad():
    prediction = model(dummy_ct, prompt)  # Output: (1, 1, 64, 128, 128)
    probabilities = torch.sigmoid(prediction)
    binary_mask = (probabilities > 0.5).cpu().numpy()

print(f"Predicted spleen mask shape: {binary_mask.shape}")

3. CLI Training & Evaluation

accelerate launch -m promptm_unet.cli train \
  --path.raw_data_dir "/path/to/totalsegmentator_raw" \
  --path.prepro_data_dir "/path/to/preprocessed_data" \
  --training.batch_size 2 \
  --training.epochs 30

# Run multi-tier evaluation
python -m promptm_unet.cli test \
  --checkpoint "./experiments/checkpoints/best_model.pth"

---

🏗️ Model Architecture

3D CT Volume (1.5mm / 3.0mm)
        │
        ▼
 ┌───────────────┐
 │ Visual Encoder│ (Residual Vision Mamba - RVM Blocks)
 └───────┬───────┘
         │
         ▼
 ┌───────────────┐        ┌──────────────────┐
 │  Bottleneck   │ ◄────► │ Text Projection  │ ◄── Frozen CLIP / BioBERT / SBERT
 └───────┬───────┘        └──────────────────┘     (512D / 768D / 384D)
         │                         ▲
         ▼                         │ (Multi-Stage Hadamard Product)
 ┌───────────────┐                 │
 │ Visual Decoder│ ────────────────┘
 └───────┬───────┘
         │
         ▼
 Deep Supervision Heads ──► High-Resolution Spleen Binary Mask

---

📜 Citation & License

This project is licensed under the Apache License 2.0.

If you find this work useful in your research, please cite our repository:

@misc{promptm_unet2026,
  title={PromptM-UNet: Efficient Text-Prompted 3D Medical Image Segmentation Using State-Space Model Mamba for Spleen},
  author={Shridhara, Ketut and Setyawan, Ivan Andika and Al-Habib, Hasanuddin},
  year={2026},
  publisher={Universitas Negeri Surabaya},
  howpublished={\url{https://github.com/kiuyha/PromptM-UNet}}
}

---

🤝 Acknowledgements

• LightM-UNet for the 3D Mamba visual backbone.
• VoxTell & SegVol for multimodal prompting inspirations.
• TotalSegmentator for CT dataset annotations.
• Research funded by LPPM Universitas Negeri Surabaya (UNESA).