nnMamba ISBI 2025 Oral

nnMamba: 3D Biomedical Image Segmentation, Classification and Landmark Detection with State Space Model

In biomedical image analysis, capturing long-range dependencies is crucial. Traditional Convolutional Neural Networks (CNNs) are limited by their local receptive fields, while Transformers, although proficient at global context integration, are computationally demanding for high-dimensional medical images.

We introduce nnMamba, a novel architecture that leverages the strengths of CNNs along with the long-range modeling capabilities of State Space Models (SSMs). Our method features the Mamba-In-Convolution with Channel-Spatial Siamese Learning (MICCSS) block that effectively models long-range voxel relationships. Additionally, we employ channel scaling and channel-sequential learning techniques to enhance performance across dense prediction and classification tasks.

Extensive experiments on seven datasets demonstrate that nnMamba outperforms current state-of-the-art methods in 3D image segmentation, classification, and landmark detection. By integrating the local representation power of CNNs with the global context processing of SSMs, nnMamba sets a new benchmark for modeling long-range dependencies in medical image analysis.

The nnMamba framework is designed for 3D biomedical tasks, focusing on dense prediction and classification. Our approach addresses the challenge of long-range modeling by harnessing the lightweight and robust capabilities of State Space Models.

Deployment

For segmentation or landmark detection task, please refer to nnMamba.py; For classification task, please refer to nnMamba4cls.py. The detailed training pipelines are available at nnunet folder for segmentation and classification folder for ADNI classification.

Checkpoints are available at:
https://drive.google.com/drive/folders/1wYHVSvSU-wGJdU62WZ0B5rfv_bZ68bJl?usp=drive_link

Methods

Architecture Overview

• Dense Prediction (Segmentation and Landmark Detection): Panels (a) and (b) illustrate the network structure.
• Classification: Panel (b) shows the network structure.
• Detailed Blocks: Panels (c), (d), and (e) provide specifics of the blocks used within the networks.

Algorithm: CSS – Channel-Spatial Siamese Learning

1. SiamSSM
   SSM with shared parameters.
2. x_flat ← input feature with shape [B, L, C]
3. x_mamba ← SiamSSM(x_flat)
4. For each dimension set d in { [1], [2], [1, 2] }:
   • x_flip ← flip(x_flat, dims = d)
   • x_mamba ← x_mamba + flip(SiamSSM(x_flip), dims = d)
5. x_mamba ← (1/4) × x_mamba

Visualization on the AMOS22 CT validation dataset: By modeling long-range dependencies, nnMamba reduces over-segmentation and under-segmentation, especially over long distances.

Results

BraTS 2023 Glioma Segmentation

Methods	WT	TC	ET	Average	WT	TC	ET	Average
DIT [Peebles et al., 2023]	93.49	90.22	84.38	89.36	4.21	5.27	13.64	7.71
UNETR [Hatamizadeh et al., 2022]	93.33	89.89	85.19	89.47	4.76	7.27	12.78	8.27
nnUNet [Isensee et al., 2021]	93.31	90.24	85.18	89.58	4.49	4.95	11.91	7.12
nnMamba	93.46	90.74	85.72	89.97	4.18	5.12	10.31	6.53

Note: The first four columns correspond to Dice scores, while the last four columns report HD95 values.

AMOS2022 Dataset

Methods	Parameters (M)	FLOPs (G)	CT-Test mDice	CT-Test mNSD	MRI-Test mDice	MRI-Test mNSD
nnUNet [Isensee et al., 2021]	31.18	680.31	89.04	78.32	67.63	59.02
nnFormer [Zhou et al., 2023]	150.14	425.78	85.61	72.48	62.92	54.06
UNETR [Hatamizadeh et al., 2022]	93.02	177.51	79.43	60.84	57.91	47.25
SwinUNetr [Hatamizadeh et al., 2021]	62.83	668.15	86.32	73.83	57.50	47.04
U-mamba [Ma et al., 2024]	40.00	792.87	87.53	75.83	74.21	64.79
nnMamba	15.55	141.14	89.63	79.73	73.98	65.13

ADNI Classification

NC vs. AD Classification

Methods	ACC	F1	AUC
ResNet [He et al., 2016]	88.40 ± 3.41	88.00 ± 2.81	94.93 ± 0.72
DenseNet [Huang et al., 2017]	87.95 ± 0.70	86.93 ± 0.87	94.86 ± 0.40
ViT [Dosovitskiy et al., 2021]	88.85 ± 1.17	87.66 ± 1.72	94.12 ± 1.29
CRATE [Yu et al., 2023]	84.69 ± 2.53	82.66 ± 3.47	91.42 ± 1.43
nnMamba	89.53 ± 0.68	88.16 ± 1.16	95.76 ± 0.18

sMCI vs. pMCI Classification

Methods	ACC	F1	AUC
ResNet [He et al., 2016]	67.96 ± 1.50	52.14 ± 1.51	74.94 ± 2.18
DenseNet [Huang et al., 2017]	73.12 ± 3.10	53.30 ± 2.99	76.31 ± 3.09
ViT [Dosovitskiy et al., 2021]	67.16 ± 3.16	51.68 ± 5.72	75.08 ± 6.88
CRATE [Yu et al., 2023]	70.63 ± 2.60	53.41 ± 2.53	76.06 ± 2.98
nnMamba	68.06 ± 4.65	53.43 ± 1.64	77.55 ± 1.29

Landmark Detection

LFC Test Set

Methods	TCD1	TCD2	HDV1	HDV2	ADV1	ADV2	Average
ResUNet [Xu et al., 2019]	1.38 ± 0.07	1.42 ± 0.10	1.46 ± 0.09	1.41 ± 0.04	1.52 ± 0.00	1.18 ± 0.05	1.39 ± 0.01
Hourglass [Newell et al., 2016]	1.40 ± 0.02	1.39 ± 0.03	1.52 ± 0.03	1.45 ± 0.04	1.47 ± 0.05	1.24 ± 0.02	1.41 ± 0.02
VitPose [Xu et al., 2022]	1.65 ± 0.01	1.73 ± 0.05	1.69 ± 0.03	1.71 ± 0.04	1.74 ± 0.04	1.38 ± 0.03	1.65 ± 0.02
SwinUnetr [Hatamizadeh et al., 2021]	1.81 ± 0.01	1.87 ± 0.03	1.82 ± 0.03	1.87 ± 0.02	1.94 ± 0.02	1.42 ± 0.03	1.79 ± 0.02
nnMamba	1.27 ± 0.01	1.40 ± 0.02	1.48 ± 0.00	1.35 ± 0.02	1.43 ± 0.01	1.14 ± 0.04	1.34 ± 0.01

FeTA Test Set

Methods	TCD1	TCD2	HDV1	HDV2	ADV1	ADV2	Average
ResUNet [Xu et al., 2019]	1.90 ± 0.30	1.46 ± 0.23	2.19 ± 0.23	1.96 ± 0.36	2.55 ± 0.48	1.73 ± 0.02	1.97 ± 0.27
Hourglass [Newell et al., 2016]	2.43 ± 0.48	1.47 ± 0.01	2.27 ± 0.04	2.10 ± 0.33	2.85 ± 0.28	1.75 ± 0.05	2.15 ± 0.12
VitPose [Xu et al., 2022]	8.46 ± 3.36	9.88 ± 0.97	16.47 ± 4.11	5.62 ± 0.98	14.46 ± 3.86	7.07 ± 3.24	10.32 ± 1.64
SwinUnetr [Hatamizadeh et al., 2021]	8.41 ± 0.87	6.50 ± 1.29	3.83 ± 0.45	4.16 ± 0.49	4.62 ± 0.39	2.44 ± 0.17	4.99 ± 0.25
nnMamba	1.70 ± 0.10	1.41 ± 0.02	1.96 ± 0.02	1.65 ± 0.03	2.20 ± 0.04	1.61 ± 0.03	1.76 ± 0.01

Citation

If you find this project useful, please consider citing us:

@inproceedings{gong2025nnmamba,  
  title={nnmamba: 3D biomedical image segmentation, classification and landmark detection with state space model},  
  author={Gong, Haifan and Kang, Luoyao and Wang, Yitao and Wang, Yihan and Wan, Xiang and Wu, Xusheng and Li, Haofeng},  
  booktitle={ISBI},  
  year={2025}  
}