1 Introduction↩︎

3D medical image segmentation is a fundamental task in medical image analysis, since accurate voxel-wise delineation of organs and lesions can directly support diagnosis, treatment planning, and disease monitoring [1]–[4]. Existing convolution-based methods have achieved strong performance, but their locality-biased operators often struggle to model long-range dependencies within volumetric data [5], [6]. Transformer-based models alleviate this limitation through self-attention, but their computational and memory costs become substantial for high-dimensional 3D medical images [7], [8]. To improve the efficiency of long-range modeling, state space models, particularly Mamba, have recently attracted increasing attention in 2D and 3D medical image segmentation [9]–[12].

Despite their linear sequence modeling capability, applying Mamba effectively to volumetric segmentation remains non-trivial. Since Mamba operates on one-dimensional sequences, 3D feature maps are usually flattened into long token sequences, making the representation sensitive to scan order and potentially disturbing volumetric spatial organization. Recent methods such as SegMamba and SegMamba-V2 address this issue by introducing tri-orientation volumetric scanning [13], [14]. However, these designs still leave several limitations. First, repeatedly scanning large 3D feature maps along multiple directions can still introduce substantial FLOPs, limiting the practical efficiency advantage of Mamba in volumetric settings. Second, directional responses are fused by fixed summation, implicitly assuming equal importance for all scan directions. Third, each directional scan is forward-only, which introduces causal asymmetry and weakens contextual modeling for voxels appearing early in the scan order. Finally, isotropic downsampling can compress the depth axis too aggressively, potentially discarding useful through-plane information in anisotropic volumes.

To address these limitations, we propose BiSegMamba, an efficient bidirectional tri-oriented Mamba for 3D medical image segmentation. The proposed model first uses a PCS to reduce redundant high-resolution computation while preserving shallow spatial details for final reconstruction. It then adopts a MSSM in the early stages to capture local anatomical patterns with low cost, and employs Bi-ToOM blocks in deeper stages to model long-range volumetric dependencies from multiple orthogonal views. In Bi-ToOM, forward and backward directional sequences are processed in a batched manner and adaptively fused with learnable branch weights, reducing causal ordering bias without introducing heavy repeated computation. In addition, an ADF module learns to weight the responses from different scan orientations instead of relying on fixed summation. As shown in Fig. 1, our model reduces FLOPs by up to 77.9% compared with SegMamba-V2 while maintaining competitive segmentation accuracy. The main contributions of this work are summarized as follows:

• We propose BiSegMamba, an efficient 3D medical image segmentation architecture that integrates PCS, MSSM, and selective deep-stage Mamba modeling to reduce redundant volumetric computation while preserving local anatomical details.

• We introduce Bi-ToOM, a bidirectional tri-oriented Mamba block that performs batched forward-backward scanning across multiple orthogonal views and uses learnable branch weighting to alleviate causal asymmetry with limited additional cost.

• We develop ADF, which learns input-dependent scan-orientation weights to aggregate directional features more flexibly than fixed summation. Extensive experiments on collected carotid CTA dataset and three public benchmarks demonstrate a favorable accuracy-efficiency trade-off against representative Mamba-based competitors.

2 Related Work↩︎

CNN- and Transformer-based Medical Image Segmentation: CNN-based encoder-decoder architectures have been the dominant paradigm for medical image segmentation. U-Net introduced skip-connected multi-scale feature fusion for dense prediction [15], while V-Net and 3D U-Net extended this idea to volumetric segmentation using 3D convolutions [5], [16]. Later methods improved feature representation through residual learning, attention gates, large-kernel convolutions, or automatic configuration, including Attention U-Net, SegResNet, UX-Net, MedNeXt, and nnU-Net [6], [17]–[20]. These methods provide strong local feature extraction, but their convolutional operators remain locality-biased and may require deep stacking or large kernels to capture long-range volumetric dependencies.

Transformer-based methods address this limitation by introducing self-attention for global context modeling. UNETR uses a ViT encoder to learn global volumetric representations [7], while TransBTS and CoTr combine convolutional feature extraction with transformer-based global modeling [21], [22]. Swin-UNet, SwinUNETR, and nnFormer further improve hierarchical representation learning through window-based or volume-based attention mechanisms [8], [23], [24]. More recent models such as TransUNet, and MISSFormer continue this direction with stronger hybrid and efficient attention designs [25], [26]. Despite their effectiveness, self-attention-based methods still introduce considerable computational and memory overhead in high-dimensional 3D medical images, motivating the search for more efficient global modeling mechanisms.

State Space Models and Vision Mamba: State space models have recently emerged as efficient alternatives to self-attention for long-sequence modeling. Mamba introduces an input-adaptive selective state space mechanism with linear complexity and hardware-aware computation, making it suitable for modeling long-range dependencies in large inputs [27]. Inspired by this, several vision-oriented variants have adapted Mamba to image understanding. Vision Mamba and VMamba introduce bidirectional or cross-scan spatial modeling strategies to process visual tokens efficiently [11], [28]. These methods demonstrate that state space models can provide effective global context modeling in vision tasks. However, most early vision Mamba models are designed for natural images and do not directly address the geometric and anisotropic characteristics of 3D medical volumes.

Mamba-based Medical Image Segmentation: Recent studies have explored Mamba-based architectures for medical image segmentation due to their ability to model long-range dependencies with lower complexity than self-attention. Mamba-UNet and VM-UNet introduce Mamba-style token mixing into U-Net-like segmentation frameworks for medical images [9], [10]. U-Mamba integrates Mamba blocks into the nnU-Net pipeline to enhance long-range dependency modeling while preserving strong convolutional priors [12]. nnMamba further investigates state space modeling for multiple 3D biomedical tasks, including segmentation, classification, and landmark detection [29]. Although these methods show the potential of Mamba in medical imaging, they do not fully address how scan directions should be organized, fused, and adapted for volumetric data.

SegMamba is among the first methods to apply Mamba-based long-range sequential modeling to 3D medical image segmentation [13]. SegMamba-V2 further improves this direction by introducing tri-oriented spatial scanning and hierarchical scale downsampling for general 3D medical segmentation [14]. However, several limitations remain: directional outputs are usually fused by fixed summation, forward-only scanning introduces causal asymmetry, and repeated multi-orientation scanning over volumetric features can still be computationally expensive. In contrast, our BiSegMamba introduces bidirectional tri-oriented Mamba modeling with learnable forward–backward fusion, adaptive directional fusion, and compact early-stage processing to improve the accuracy–efficiency trade-off in 3D medical image segmentation.

3 Methodology↩︎

3.1 Overall Architecture↩︎

Fig. 2 shows the overall architecture of BiSegMamba. The model is built around a compact-to-detail design: expensive hierarchical reasoning is performed in a compact latent space, while fine spatial cues are preserved through a lightweight high-resolution pathway for final reconstruction. This design is motivated by the observation that directly applying Mamba-based processing to dense 3D feature maps is computationally expensive, whereas aggressive early downsampling may remove boundary information that is important for thin vessels, plaques, small organs, and tumor margins. To balance these two aspects, the input volume is first processed by PCS, which produces a compact encoder input and a shallow high-resolution feature. The compact representation is passed through a four-stage encoder following a local-to-global hierarchy. The first two stages use MSSM blocks for efficient local anatomical representation, while the deeper stages use Bi-ToOM block for long-range volumetric context modeling from orthogonal views. The decoder progressively restores spatial resolution through skip fusion. Finally, Reconstruction Head (RCH) combines the decoded feature with the preserved shallow feature from PCS, allowing the model to recover fine boundary details while keeping most computation in the compact pathway. This architecture separates three complementary roles: PCS reduces redundant high-resolution computation, MSSM strengthens local spatial representation before global sequence modeling, and Bi-ToOM provides efficient long-range context aggregation at semantic resolutions. The following subsections describe these components in detail.

3.2 Progressive Compacting Stem↩︎

PCS converts the input volume into two complementary representations: a compact latent tensor \(\mathbf{x}_0\) for encoder processing and a shallow high-resolution feature \(\mathbf{s}\) for final reconstruction. Unlike a conventional patch embedding layer that only reduces spatial resolution, PCS explicitly preserves early structural cues so that latent-space computation does not discard fine boundary information. For a given input volume \(\mathbf{x}\in\mathbb{R}^{B\times C_{\mathrm{in}}\times D\times H\times W}\), PCS is formulated as \[\begin{align} \mathbf{s} &= f_2(f_1(\mathbf{x})),\\ \mathbf{u} &= g_1(\mathbf{s}),\\ \mathbf{x}_0 &= g_2\!\left(\mathbf{u} + r(\mathbf{u})\right), \end{align}\] where \(f_1(\cdot)\) and \(f_2(\cdot)\) are shallow feature extractors, \(g_1(\cdot)\) and \(g_2(\cdot)\) are progressive reduction operators, and \(r(\cdot)\) is a lightweight refinement branch. The compact representation \(\mathbf{x}_0\) is passed to the main encoder, while \(\mathbf{s}\) is retained as a high-resolution skip feature and fused only in the final reconstruction head.

Methodologically, PCS decouples representation learning from spatial recovery: the encoder operates on \(\mathbf{x}_0\) to reduce the cost of subsequent stages, while \(\mathbf{s}\) provides a direct path for recovering thin structures, boundary transitions, and local intensity discontinuities. For anisotropic data, PCS uses depth-preserving early reduction so that through-plane information is compacted more conservatively than in-plane information; for near-isotropic inputs, the same formulation is instantiated with symmetric spatial reduction.

3.3 Multi-Scale Spatial Mixer↩︎

The first two encoder stages operate on the compact latent representation produced by PCS, where local anatomical cues are still important but dense sequence modeling remains computationally expensive. We therefore use MSSM as a lightweight local mixing module before Mamba-based global modeling. Let \(\mathbf{u}\in\mathbb{R}^{B\times C\times D\times H\times W}\) denote the input feature tensor to an MSSM block; for the first encoder stage, \(\mathbf{u}\) is initialized from the PCS output \(\mathbf{x}_0\). MSSM applies parallel depthwise convolution branches with complementary receptive fields: \[\begin{align} \mathbf{m}_1 &= \phi(\mathrm{IN}(\mathrm{DW}_{3\times3\times3}(\mathbf{u}))),\\ \mathbf{m}_2 &= \phi(\mathrm{IN}(\mathrm{DW}_{5\times5\times5}(\mathbf{u}))),\\ \mathbf{m}_3 &= \phi(\mathrm{IN}(\mathrm{DW}^{\mathrm{dil}=2}_{3\times3\times3}(\mathbf{u}))), \end{align}\] where \(\phi(\cdot)\) denotes GELU and \(\mathrm{IN}\) denotes instance normalization. For anisotropic inputs, an additional in-plane branch is used to strengthen local context modeling without excessive through-plane smoothing: \[\mathbf{m}_4 = \phi(\mathrm{IN}(\mathrm{DW}_{1\times5\times5}(\mathbf{u}))).\] The branch outputs are aggregated, refined by a depthwise convolution, recalibrated by squeeze-and-excitation, and added to the input with a learnable residual scale: \[\mathrm{MSSM}(\mathbf{u}) = \mathbf{u} + \boldsymbol{\gamma}\odot \mathrm{SE}\!\left( \phi(\mathrm{IN}(\mathrm{DW}_{3\times3\times3}(\sum_i \mathbf{m}_i))) \right),\] where the summation is taken over the active branches. The anisotropic variant is used in the first stage, where the depth resolution is still preserved, while the second stage uses the isotropic variant. MSSM thus provides efficient multi-scale local representation before long-range Mamba-based modeling.

3.4 Bidirectional Tri-Oriented Ortho Mamba Block↩︎

The deeper encoder stages are responsible for modeling long-range dependencies over compact volumetric representations. Inspired by recent Mamba-based 3D segmentation methods [13], [14], we represent a 3D feature tensor using three orthogonal scan orientations, allowing contextual information to be collected from different spatial orderings instead of a single flattening path. However, existing tri-oriented designs mainly rely on forward-only scanning and fixed directional aggregation, which can introduce scan-order bias and limit the adaptability of directional context fusion. In response, we introduce the Bi-ToOM block, as illustrated in Fig. 3. Bi-ToOM differs from previous tri-oriented Mamba blocks in two key aspects. First, each directional sequence is processed in both forward and backward order, allowing each voxel to receive context from both preceding and succeeding positions in the flattened sequence. Second, all three directional sequences and their reversed counterparts are concatenated along the batch dimension and processed by a single Mamba operator, avoiding six independent Mamba calls. The resulting forward and backward responses are then fused with learnable channel-wise weights, enabling the network to adaptively balance past and future context within each scan orientation.

For a given input tensor \(\mathbf{u}\in\mathbb{R}^{B\times C\times D\times H\times W}\), we first construct three orthogonal directional views, \(\mathbf{u}^{(d)}=\mathbf{u}, \mathbf{u}^{(w)}=\Pi_w(\mathbf{u}), \mathbf{u}^{(h)}=\Pi_h(\mathbf{u}),\) where \(\Pi_w(\cdot)\) and \(\Pi_h(\cdot)\) are spatial permutations that make width and height the leading scan axis, respectively. Each directional tensor is then flattened into a sequence of length \(N=DHW\):

\[\mathbf{s}^{(i)}=\mathrm{Flat}\!\left(\mathbf{u}^{(i)}\right)\in\mathbb{R}^{B\times N\times C}, \qquad i\in\{d,w,h\}.\] To enable bidirectional context aggregation, we additionally form reversed sequences \[\bar{\mathbf{s}}^{(i)}=\mathrm{Rev}\!\left(\mathbf{s}^{(i)}\right), \qquad i\in\{d,w,h\}.\] Rather than processing each branch independently, all six sequences are concatenated along the batch dimension and passed through a single normalized Mamba operator: \[\mathbf{S}=\mathrm{Cat}\!\left[ \mathbf{s}^{(d)},\mathbf{s}^{(w)},\mathbf{s}^{(h)}, \bar{\mathbf{s}}^{(d)},\bar{\mathbf{s}}^{(w)},\bar{\mathbf{s}}^{(h)} \right],\] \[\mathbf{O}=\mathrm{Mamba}(\mathrm{LN}(\mathbf{S})).\] This batched realization preserves the efficiency of a single Mamba kernel call while allowing every directional view to be modeled in both forward and backward order. The output tensor \(\mathbf{O}\) is then split back into six sequence groups, \[(\mathbf{o}^{(d)}_f,\mathbf{o}^{(w)}_f,\mathbf{o}^{(h)}_f, \mathbf{o}^{(d)}_b,\mathbf{o}^{(w)}_b,\mathbf{o}^{(h)}_b),\] where the \(f\) denotes the forward branch and \(b\) denotes the backward branch. Each sequence is reshaped to its corresponding 3D layout, and the backward output is reversed back to the original spatial order. For each direction, the forward and backward responses are combined through learnable channel-wise weights: \[\hat{\mathbf{u}}^{(i)}= \alpha^{(i)}_f\odot \mathbf{u}^{(i)}_f + \alpha^{(i)}_b\odot \mathbf{u}^{(i)}_b, \qquad i\in\{d,w,h\}, \label{eq:tensors}\tag{1}\] where \((\alpha^{(i)}_f,\alpha^{(i)}_b)\) are obtained by a softmax over a pair of learnable logits for each direction and channel. In this way, the network can adaptively control the relative contribution of past and future context within each scan orientation. After bidirectional fusion, the three directional tensors are restored to the common spatial layout and passed to ADF, which learns the reliability of each scan orientation before producing the final residual output.

3.5 Adaptive Directional Fusion↩︎

After bidirectional scanning, the three orthogonal views provide complementary but not equally reliable contextual information. A simple summation assigns the same importance to all scan orientations, which may introduce less informative directional responses into the fused representation. This is suboptimal because the most useful direction can vary with anatomical shape, local structure, and imaging geometry. To address this, we introduce ADF, a lightweight fusion module that learns input-adaptive directional weights before aggregation as shown in Fig. 3.

For the three directional tensors \(\mathbf{u}^{(d)}, \mathbf{u}^{(w)}, \mathbf{u}^{(h)} \in \mathbb{R}^{B\times C\times D\times H\times W}\) from Eq. 1 , ADF first concatenates them along the channel dimension and applies a lightweight gating network to predict channel-wise directional logits: \[\boldsymbol{\ell}= \mathcal{G}\!\left( [\mathbf{u}^{(d)};\mathbf{u}^{(w)};\mathbf{u}^{(h)}] \right) \in\mathbb{R}^{B\times 3C\times1\times1\times1}.\] The logits are reshaped to \(\mathbb{R}^{B\times3\times C\times1\times1\times1}\) and normalized across the three scan orientations: \[\boldsymbol{\alpha}=\mathrm{softmax}(\boldsymbol{\ell}, \mathrm{dim}=1).\]

The normalized weights are applied to the three directional features, which are then concatenated and projected by a \(1\times1\times1\) mixing layer: \[\mathbf{v}= \mathcal{M}\!\left( \mathrm{reshape}\!\left( \boldsymbol{\alpha}\odot \mathrm{stack}(\mathbf{u}^{(d)},\mathbf{u}^{(w)},\mathbf{u}^{(h)}) \right)\right).\] Finally, the fused response is optionally refined by squeeze-and-excitation and modulated by a learnable residual scale \(\boldsymbol{\gamma}\): \[\mathrm{ADF}(\mathbf{u}^{(d)},\mathbf{u}^{(w)},\mathbf{u}^{(h)}) = \boldsymbol{\gamma}\odot \mathrm{SE}(\mathbf{v}).\]

ADF therefore replaces fixed directional summation with input-adaptive channel-wise aggregation, allowing BiSegMamba to emphasize the most informative scan orientations for each volumetric representation.

3.6 Decoder and Reconstruction Head↩︎

The decoder adopts a UNETR-style hierarchical reconstruction pathway to progressively recover spatial resolution from the encoder feature pyramid. Given the encoder outputs, the decoder performs three upsampling stages, each fused with the corresponding encoder skip feature. This skip-fusion design transfers multi-scale semantic information from the compact encoder pathway to the reconstruction process. The final prediction is RCH, which maps the last decoder feature back to the original image space and reintroduces the shallow high-resolution feature preserved by PCS. Specifically, \[\mathbf{y}=\mathcal{H}(\mathbf{d}_1,\mathbf{s}),\qquad \hat{\mathbf{p}}=\mathrm{Conv}_{1\times1\times1}(\mathbf{y}),\] where \(\mathbf{d}_1\) is the final decoder feature, \(\mathbf{s}\) is the shallow feature from PCS, and \(\mathcal{H}(\cdot)\) denotes the reconstruction head. By fusing \(\mathbf{s}\) only at the final stage, RCH provides high-resolution structural guidance without repeatedly processing dense features throughout the encoder.

For anisotropic inputs, the first transposed convolution in \(\mathcal{H}\) uses kernel and stride \((1,2,2)\), restoring in-plane resolution without unnecessary expansion along the depth axis. This geometry-aware reconstruction preserves through-plane consistency while recovering fine boundary details for voxel-wise prediction.

4 Experiments↩︎

4.1 Datasets↩︎

4.1.1 Carotid CTA Dataset↩︎

We collected a carotid dataset consisting of 115 retrospectively acquired and fully anonymized head-and-neck CTA volumes from Peking University Third Hospital between January 2020 and June 2024. The data were split at the patient level into 80 training, 18 validation, and 17 test volumes. Scans were acquired using GE Revolution CT and Siemens SOMATOM Force scanners, with an in-plane size of \(512\times512\), axial spacing of 0.488/0.500/0.625 mm, and 312–945 slices per volume. Ground-truth masks were annotated in 3D Slicer under expert-doctor guidance. Each volume was first initialized by grayscale-threshold-based coarse vascular localization and then refined slice-by-slice through manual verification and morphological correction. The annotation range covered major head-and-neck arterial structures from the lower cervical/aortic-arch region to the intracranial internal carotid artery C6 segment.

4.1.2 Public Benchmarks↩︎

We further evaluate the proposed method on three public benchmarks: ACDC, BraTS2023, and AMOS-CT. The ACDC dataset is a cardiac MRI segmentation benchmark containing 100 patient volumes with annotations for the right ventricle (RV), myocardium (Myo), and left ventricle (LV) [30]. BraTS2023 is a multi-parametric brain tumor MRI segmentation benchmark, where each case includes T1, T1Gd, T2, and T2-FLAIR modalities with labels for whole tumor (WT), tumor core (TC), and enhancing tumor (ET) [14], [31]. AMOS-CT is used for abdominal multi-organ segmentation and provides voxel-level annotations for 15 organs, including large organs and small anatomically variable structures such as the adrenal glands, pancreas, esophagus, and duodenum [32]. These benchmarks provide complementary evaluation scenarios across cardiac MRI, brain tumor MRI, and abdominal CT segmentation.

4.2 Implementation Details↩︎

All experiments were implemented in PyTorch and conducted on a single NVIDIA RTX 4090 GPU. Dataset-specific training protocols were adopted for fair comparison with representative baselines. For the carotid CTA dataset, we followed an nnFormer/nnU-Net-style pipeline with carotid-specific modifications and used a crop size of \(128\times128\times128\). For BraTS2023, we followed the SegMamba-V2 setting with the same crop size. For AMOS-CT and ACDC, the crop sizes were set to \(64\times160\times160\) and \(16\times160\times160\), respectively, following their reference experimental settings.

The proposed model was trained end-to-end on all datasets using stochastic gradient descent (SGD) with Nesterov momentum of 0.99, an initial learning rate of 0.01 for the public benchmarks, and polynomial learning-rate decay. For the carotid CTA dataset, we used a lower initial learning rate of 0.005 for stable plaque learning, together with foreground oversampling and 3D data augmentation. The carotid CTA task was formulated as a three-class segmentation problem, including background, vessel, and plaque, with foreground computed as the union of vessel and plaque for evaluation. Test-time mirroring was disabled for carotid CTA to preserve clinically meaningful left-right vascular anatomy and plaque localization. For the public benchmarks, preprocessing, augmentation, and inference settings were kept consistent with the corresponding reference protocols. During inference, 3D sliding-window prediction was used, and the outputs were converted to voxel-wise masks in the original label space of each dataset.

4.3 Evaluation Metrics↩︎

We evaluate segmentation performance using the Dice similarity coefficient (DSC) and the 95% Hausdorff distance (HD95), which are two of the most commonly used metrics in volumetric medical image segmentation. DSC measures the overlap between the predicted mask and the ground-truth annotation, while HD95 evaluates boundary quality by computing the 95th percentile of the bidirectional surface distance between prediction and ground truth. In this work, DSC is used as the primary overlap-based metric, and HD95 is used as a complementary boundary-sensitive metric.

4.4 Quantitative Comparison Against State-of-the-Art Methods↩︎

4.4.1 Carotid Dataset↩︎

Table 1 reports the quantitative comparison on the carotid CTA dataset. BiSegMamba achieves the best results across vessel, plaque, and foreground segmentation. For vessel segmentation, it obtains a Dice score of 91.63% and an HD95 of 5.63 mm, outperforming the strongest baseline, nnU-Net, in both overlap and boundary accuracy. The gain is more evident for plaque segmentation, where BiSegMamba improves the Dice score from 45.38% with U-Mamba to 60.99%, and reduces HD95 from 32.04 mm with nnU-Net to 15.13 mm. This indicates better localization of small and irregular plaque regions. For foreground segmentation, BiSegMamba also achieves the best Dice and HD95, reaching 96.38% and 4.42 mm, respectively. These results show that the proposed model is effective for both elongated carotid vessel delineation and fine-grained plaque segmentation.

4.2pt

4.4.2 BraTS2023↩︎

Table 2 summarizes the quantitative results on BraTS2023. BiSegMamba achieves the highest average Dice score of 91.63%, slightly outperforming SegMamba-V2 while surpassing the compared CNN- and Transformer-based baselines. At the region level, it obtains the best Dice scores on TC and ET and the second-best Dice on WT, indicating strong performance on challenging tumor subregions. For boundary accuracy, BiSegMamba achieves the best HD95 on ET and remains competitive in average HD95. Although SegMamba-V2 obtains a slightly lower average HD95, the difference is small, while BiSegMamba provides the best average overlap accuracy. Importantly, this performance is achieved with lower complexity than SegMamba-V2, reducing parameters from 138.77M to 47.38M and FLOPs from 1853.19G to 410.28G. These results demonstrate a favorable accuracy-efficiency trade-off on multi-modal brain tumor segmentation.

4.2pt

4.4.3 ACDC↩︎

Table 3 reports the DSC comparison on the ACDC dataset. BiSegMamba achieves the best average DSC of 92.57% and obtains the highest scores on all three cardiac structures, including RV, Myo, and LV. Compared with strong CNN- and Transformer-based baselines such as nnU-Net and nnFormer, the proposed method shows consistent improvements in region overlap accuracy. Since SegMamba-V2 did not report ACDC results, we reproduced it under the same experimental setting for a controlled comparison. BiSegMamba improves the average DSC from 90.79% to 92.57%, with consistent gains across all cardiac structures. The reported \(p\)-value further confirms that the improvement over SegMamba-V2\(^\dagger\) is statistically significant. These results indicate that the proposed design transfers effectively to cardiac MRI segmentation.

5.0pt

4.4.4 AMOS-CT↩︎

Table 4 reports the Dice comparison on the AMOS-CT validation set. BiSegMamba achieves the highest mDice of 89.03%, outperforming both CNN/Transformer-based baselines and recent Mamba-based methods. Compared with the strongest non-Mamba baseline, UNet, it improves mDice from 88.87% to 89.03%. The gain is larger over recent Mamba-based models, improving over Mamba-HoME from 86.30% to 89.03%. At the organ level, BiSegMamba obtains the best Dice score on 11 out of 15 structures, including several challenging small or anatomically variable organs such as the esophagus, pancreas, right adrenal gland, and duodenum. Although it is not the best on gallbladder, right kidney, left adrenal gland, and prostate/uterus, it remains competitive on these categories while achieving the best overall mDice. These results demonstrate strong and balanced multi-organ segmentation performance on AMOS-CT.

3.0pt

4.5 Ablation Study↩︎

Table 5 reports the ablation study on the ACDC dataset. Starting from SegMamba-V2, replacing the early-stage DWConv blocks with MSSM improves the average DSC from 90.79% to 91.30% and reduces HD95 from 1.37 to 1.25, while also lowering FLOPs. This indicates that multi-scale local spatial mixing is more effective and efficient than the original early convolutional design. Introducing PCS further improves the DSC to 92.12% and reduces FLOPs from 290.04G to 130.73G, confirming the benefit of compact latent-space processing. Replacing HSDownsampling with patch merging provides an additional gain, reaching 92.20% DSC and 1.16 HD95. The contribution of bidirectional Mamba modeling is shown by the Bi-ToOM variant, which improves the DSC to 92.35% and reduces HD95 to 1.11. Adding ADF further increases the DSC to 92.53% and lowers HD95 to 1.08, demonstrating the advantage of adaptive directional aggregation over fixed directional summation. Among all variants, BiSegMamba-ACDC\(^\ast\) achieves the best ACDC performance, with 92.57% DSC and 1.06 HD95. This accuracy-oriented configuration keeps GSC-based local conditioning before the deepest ToOM block, which is beneficial for cardiac MRI segmentation. In contrast, the default BiSegMamba replaces this stage-4 GSC with MSSM, reducing the parameter count to 57.53M and FLOPs to 146.09G with only a small performance decrease. Therefore, BiSegMamba-ACDC\(^\ast\) is used for ACDC, while BiSegMamba is used as the default lightweight configuration for the other datasets.

3.4pt

4.6 Qualitative Visualization↩︎

Fig. 4, 5, 6 and 7 show qualitative comparisons on the carotid CTA dataset, ACDC, BraTS2023, and AMOS-CT, respectively. Overall, the visual results are consistent with the quantitative findings and show that the proposed BiSegMamba produces more complete and anatomically coherent segmentation masks across different imaging modalities and anatomical targets. On the carotid CTA dataset, BiSegMamba better preserves the vessel lumen structure and captures small plaque regions with fewer fragmented predictions. In comparison, several baseline methods either miss small plaque regions or introduce false-positive responses around nearby tissues. This indicates that the proposed bidirectional tri-oriented modeling is beneficial for elongated vascular structures and small lesion-like targets. For ACDC cardiac MRI, the proposed method generates more consistent RV, myocardium, and LV boundaries across different slices. Compared with nnU-Net, nnFormer, UNETR, and SegMamba-V2, BiSegMamba shows fewer boundary discontinuities and better preserves the thin myocardial ring, especially in regions where adjacent cardiac structures have similar intensity. On BraTS2023, BiSegMamba produces tumor masks that are closer to the ground truth, particularly around irregular tumor boundaries and enhancing tumor regions. Compared with SegMamba-V2, the proposed method better maintains the spatial relationship among necrotic tumor, edema, and enhancing tumor regions, reducing local under-segmentation and boundary inconsistency. For AMOS-CT multi-organ segmentation, the proposed method provides more stable organ delineation across large organs and small anatomical structures. Compared with SegMamba, BiSegMamba better separates neighboring abdominal organs and reduces missing or fragmented predictions for small structures. These qualitative results suggest that the proposed architecture improves both global structural consistency and local boundary precision in challenging 3D medical image segmentation tasks.

5 Conclusion↩︎

In this paper, we presented BiSegMamba, an efficient bidirectional tri-oriented Mamba framework for 3D medical image segmentation. The proposed architecture combines a progressive compacting stem, multi-scale spatial mixing, bidirectional tri-oriented state-space modeling, and adaptive directional fusion to improve volumetric context modeling while maintaining computational efficiency. Unlike previous Mamba-based segmentation designs that rely on forward-only directional scanning and fixed directional aggregation, BiSegMamba enables more balanced contextual propagation and adaptive fusion of directional features. Extensive experiments on a carotid dataset and three public benchmarks, including ACDC, BraTS2023, and AMOS-CT, demonstrate the effectiveness and generalizability of the proposed method. BiSegMamba achieves strong performance across vascular, cardiac, brain tumor, and abdominal multi-organ segmentation tasks, with particularly clear advantages in challenging settings involving elongated structures, small lesion regions, and complex anatomical boundaries. The ablation study further validates the contribution of each core component, including MSSM, PCS, Bi-ToOM, and adaptive directional fusion. Overall, the proposed method provides a favorable accuracy-efficiency trade-off for 3D medical image segmentation. In future work, we will further evaluate the framework on larger multi-center clinical datasets and explore its extension to uncertainty-aware and foundation-model-assisted volumetric segmentation.

References↩︎

---

1. This work was partially supported by the National Natural Science Foundation of China (62572033, 62176016, 72274127), Beijing Municipal Science and Technology Program (Z251100003625009), Guizhou Province Science and Technology Project (Qiankehe[2024] General 058), Haidian Innovation and Translation Program of Peking University Third Hospital (HDCXZHKC2023203), and the Digital Technology-Empowered Urban and Park Carbon Emission Decision Support Project.↩︎

2. School of Computer Science and Engineering, Beihang University. State Key Laboratory of Virtual Reality Technology and Systems, Beihang University (bakhtzada@buaa.edu.cn).↩︎

3. Corresponding author: Chao Tong (tongchao@buaa.edu.cn).↩︎