Research on segmentation method of elderly cardiovascular disease feature images based on artificial intelligence multi-scale feature fusion

Abstract

Accurate segmentation of medical images is crucial for diagnosing and treating cardiovascular diseases in the elderly. However, these images often suffer from blurred boundaries and low contrast due to complex lesions such as calcification and plaque, challenging existing methods to simultaneously capture global context and preserve local details. To address this, we propose CTM-Net, a collaborative framework integrating convolutional neural networks (CNNs), transformers, and multilayer perceptrons (MLPs). The CNN encoder extracts hierarchical local features, a Transformer module at the bottleneck captures long-range dependencies, and a lightweight MLP-based decoder with a novel Spatial-Channel MLP (SC-MLP) block performs efficient upsampling and pixel-level classification. Multi-scale feature fusion is achieved via skip connections. Experiments on three public cardiovascular datasets (ASOCA, Cardiac-MRI, Sunnybrook) demonstrate that our method significantly outperforms mainstream models like U-Net, TransUNet, and nnUNet in key metrics (e.g., 82.1% Dice on ASOCA vs. 81.5% for nnUNet, p < 0.05), while maintaining superior computational efficiency. Systematic ablation studies validate the synergistic design. This framework offers a promising pathway for complex cardiovascular image segmentation.

Introduction

With the aggravation of global population aging, cardiovascular disease has become the main factor threatening the health of the elderly, and its early diagnosis and precise treatment have great clinical and social significance¹. Medical imaging techniques, such as coronary computed tomography angiography (CCTA) and cardiac magnetic resonance imaging (MRI), can non-destructively reveal the structure and function of the cardiovascular system, which is the core basis for clinical diagnosis². In this context, accurate and automatic image segmentation of key structures, such as heart, blood vessels and plaques has become an indispensable step for quantitative analysis, surgical planning and prognosis evaluation.

In recent years, deep learning technology, especially fully convolutional neural network (FCN) and its representative model U-Net, has greatly promoted the development of medical image segmentation^3,4. These convolutional neural network (CNN) based methods can effectively extract hierarchical features from local receptive fields, and have achieved remarkable success in numerous tasks. However, the inherent locality property of CNNS limits their ability to model long-range dependencies in images⁵. Recently, MLP-based models, such as MLP-Mixer¹¹ and ConvNeXt¹², have regained attention due to their structural simplicity and competitive performance, offering efficient alternatives for dense prediction tasks. This raises an intriguing question: can we construct a collaborative framework where CNN, Transformer, and MLP each perform specialized roles—local feature extraction, global context modeling, and efficient upsampling—to systematically tackle the challenges of cardiovascular image segmentation?

To overcome these limitations, researchers have introduced Transformer architectures, which excel at capturing global context via self-attention, into computer vision⁶. Vision Transformer (ViT) and its variants have demonstrated powerful performance⁷, leading to hybrid models like TransUNet⁸ that combine CNNs and Transformers. However, such models face challenges: applying global self-attention on high-resolution features is computationally expensive, and many still use traditional convolutional decoders with room for improvement in efficiency and performance.

Recently, MLP-based models, such as MLP-Mixer⁹ and ConvNeXt¹⁰, have regained attention due to their structural simplicity and competitive performance, offering efficient alternatives for dense prediction tasks. This raises the question: can we build a collaborative framework where CNN, Transformer, and MLP each perform their specialized roles—local feature extraction, global context modeling, and efficient upsampling prediction—to systematically address the core challenges of cardiovascular image segmentation.

In this article, we propose a novel collaborative segmentation framework called CTM-Net. Our key contributions are:

A Systematic CNN-Transformer-MLP Co-segmentation Framework: We design a synergistic architecture where a CNN encoder ensures detailed local feature extraction, a lightweight Transformer module at the bottleneck captures global context, and a novel MLP-based decoder enables efficient feature fusion and upsampling.

A Novel Lightweight SC-MLP Decoder: We introduce a Spatial-Channel MLP (SC-MLP) block that utilizes pixel shuffling and separate channel/spatial information processing. This achieves more accurate pixel-wise prediction with significantly fewer parameters (42% reduction) and faster inference (1.7× speedup) compared to traditional convolutional decoders.

Comprehensive Validation: We provide extensive experiments on multiple public datasets, including statistical significance testing, ablation studies, and comparisons with recent state-of-the-art models, demonstrating the effectiveness and efficiency of our approach.

Review literature

Segmentation method based on the convolutional neural network

CNNs have established a fundamental position in medical image segmentation. U-Net¹¹ constructs the classical encoder-decoder paradigm with skip connections for multi-scale feature fusion, which is particularly suitable for medical images with blurred boundaries. However, its limited receptive field hinders global context capture, especially in complex cardiovascular images where local lesion areas need to be understood in the context of overall vessel direction. Attention U-Net¹² introduces attention gates to enhance relevant features but does not fundamentally expand the receptive field. U-Net++¹³ optimizes feature fusion with nested dense skip connections, showing improved boundary segmentation but remaining limited by CNN locality. ResUNet++¹⁴ combines residual connections and attention, demonstrating the potential of carefully designed CNN modules, yet still constrained by local receptive fields.

These studies highlight the inherent limitation of pure CNN architectures in global modeling. This motivates our introduction of a Transformer module to explicitly model long-range dependencies.

Segmentation methods based on the transformer and hybrid architecture

To overcome CNN limitations, researchers have explored Transformers in vision. The success of these hybrids validates the importance of global context. Recent advances in Transformer-based architectures for visual understanding, such as those proposed by Sharma et al.¹⁵ for image captioning and summarization, further highlight the versatility of self-attention mechanisms in modeling complex visual-semantic relationships. TransUNet¹⁶ is a key breakthrough, using a CNN-Transformer hybrid encoder. However, its decoder still relies on traditional convolutional upsampling, failing to fully leverage the global features from the Transformer. Swin-Unet ¹⁷ adopts a hierarchical Transformer with sliding windows, reducing complexity but facing training stability issues on smaller datasets. nnFormer ¹⁸ integrates local and global features but has high model complexity.

The success of these hybrids validates the importance of global context. However, their decoder designs point to a need for improvement, which is the core motivation for our lightweight MLP-based decoder.

Multi-layer perceptron based segmentation methods

In recent years, multi-layer perceptron-based architectures have gained renewed attention due to their computational efficiency, which provides new ideas for medical image segmentation. The MLP-Mixer shows the potential of the full MLP architecture. This model achieves effective feature interaction through a simple combination of channel MLP and token MLP, and shows comparable performance to CNN and Transformer in image classification tasks¹⁹. However, it was originally designed for classification tasks and lacks a multi-scale feature fusion mechanism suitable for dense prediction tasks. UNeXt proposed by Song et al.²⁰, introduces MLP into the field of medical image segmentation and adopts a minimalist encoder-decoder design to achieve fast inference with a very low number of parameters. This work demonstrates the value of MLP in constructing lightweight segmentation models, but for the pursuit of efficiency, its accuracy in complex structure segmentation tasks is bottleneck-bound.

SegMLP proposed by Guo et al.²¹, designed an MLP architecture specifically for semantic segmentation tasks, and achieved good segmentation results through feature transformation of two dimensions, i.e., space and channel. This method improves the segmentation accuracy while maintaining high efficiency, but its detail preservation ability in complex medical images still needs to be strengthened. MLP-segnet proposed by Yu et al.²² explores the application of MLP in multi-modal medical image segmentation and shows the adaptability of MLP architecture in dealing with complex medical data. However, its performance still has room for improvement in single-model fine structure segmentation tasks.

The efficiency of MLP models inspires our design. We do not seek to replace CNNs or Transformers but to leverage their strengths synergistically. We draw on MLP’s efficient feature transformation to design a specialized lightweight decoder, while relying on the CNN-Transformer encoder for robust feature extraction, achieving an optimal accuracy-efficiency balance.

There is a clear trend of technology fusion in medical image segmentation. Our proposed CTM-Net aims for deep collaboration of the three architectures, not simple superposition, through carefully designed structural division of labor.

Research methodology

The proposed CTM-Net is a deep learning framework specifically designed for cardiovascular image segmentation. The overall network structure is shown in Fig. 1, consisting of three core components: a CNN-based hierarchical feature encoder, a Transformer-based global context modeling module, and an efficient MLP-based feature decoder.

Fig. 1.

CNN encoder model diagram.

CNN-based hierarchical feature encoder

In the feature extraction stage, we use EfficientNet-B4 pre-trained by ImageNet as the base network. This choice is based on many considerations. On the one hand, EfficientNet achieves a balance between depth, width and resolution through a composite scaling method, which can effectively control the computational complexity while ensuring the quality of features ²⁴. On the other hand, the pre-trained weights provide a good parameter initialization, which helps the model to converge quickly. Finally, its hierarchical structure is naturally suitable for extracting multi-scale features, which is crucial for dealing with cardiovascular structures of different sizes and shapes (Fig. 1).

Given an input cardiovascular image I∈R^(H×W×3), where, H-represents the image height, W-represents the image width, and 3-represents the three RGB color channels. We extract multi-scale features through CNN encoder. The feature extraction process can be formalized as follows:

Where, the spatial resolution of each feature map F₁ is 1/2^i of the input, and the corresponding number of channels is 24, 32, 56, 112 and 179, respectively. The selection of these channels follows the compound scaling principle of EfficientNet-B4²³, which achieves the balance between computational complexity and feature expression ability.

In practice, we specifically select the middle three levels of features to form the feature pyramid.

Here, H/8 × W/8 represents the spatial size of the feature map and 56 represents the number of feature channels. This multi-scale feature design can capture rich information from local details to global semantics. The shallow features C₃ retain rich spatial detail information, which is conducive to accurate boundary location. Mid-level features C₄ balance the detail and semantic information. The deep feature C₅ contains stronger semantic information and helps to understand the overall structure.

Transformer-based global context modeling module

To overcome the limitations of traditional CNNS in modeling long-range dependencies, we introduce the Transformer module on the deepest level feature C₅ output from the encoder. The core insight of this design is that the understanding of cardiovascular structures often requires global context information. For example, to determine the degree of stenosis in a certain vascular segment requires reference to the direction of the whole vessel, and to identify the type of plaque requires reference to the characteristics of the surrounding tissue ²⁵. The self-attention mechanism of Transformer can establish a correlation between any two locations in the image, which meets this requirement.

We first project C₅ to the Transformer embedding dimension (d_model = 192) using a 1 × 1 convolution: C₅_proj = Conv₁×₁(C₅) ∈ R^{H/32×W/32 × 192}. This projection aligns the feature channels with the Transformer’s expected input dimension.

This feature map is then reshaped into a sequence:

To preserve spatial information, we add a learnable positional encoding:

The Transformer encoder is stacked with L = 6 identical layers. Each layer contains two core submodules: multi-head self-attention (MSA) and a feed-forward network (FFN). With h = 8 attention heads, the head dimension is d_k = d_model / h = 24.

The FFN is a two-layer MLP with GELU activation:

Residual connections and layer normalization are applied after each sublayer:

This design helps to alleviate the vanishing gradient problem in deep networks. By stacking multi-layer transformers, the model is able to establish a multi-level feature representation from local to global, which fully captures the complex structural relationships in cardiovascular images (Fig. 2).

Fig. 2.

Diagram of the transformer-based global context modeling module.

Efficient MLP-based Feature decoder

The decoder undertakes the important task of restoring high-level semantic features into the original resolution segmentation map in the segmentation network. The traditional convolution-based decoder has the problem of many parameters and complex calculation. Inspired by the recent research on MLP architecture, we creatively design a lightweight decoder based on spatial-channel MLP (SC-MLP). This design not only significantly reduces the number of parameters, but also maintains excellent feature reconstruction ability (Fig. 3).

Fig. 3.

Efficient feature decoder based on MLP.

Feature Upsampling and fusion strategy

The decoder adopts a progressive upsampling scheme, which is in line with the human visual system’s processing pattern of gradual refinement from course to fine. Each stage contains a carefully designed processing flow, which ensures that multi-scale information can be effectively fused while magnifying the feature map. The upsampling process can be formalized as follows.

Where, PS stands for pixel shuffling operation, which is an efficient upsampling method that achieves resolution improvement by rearranging the information of the channel dimension, avoiding the checker board effect that may be brought by transposed convolution. C_i is the encoder feature at the corresponding scale, which provides valuable spatial detail information through skip connections.

The advantage of this progressive fusion strategy is that at each upsampling stage, the high-level semantic features are fused with the corresponding low-level detailed features in time, which not only maintains the semantic consistency, but also recovers the spatial accuracy. This multi-scale feature fusion is particularly important when dealing with fine structures, such as cardiovascular boundaries.

SC-MLP block design

The SC-MLP block is the core innovative module we propose, whose design is inspired by modern MLP architectures but specifically optimized for the medical image segmentation task. The module adopts a divide-and-conquer strategy to deal with the information interaction in two dimensions of space and channel respectively, which ensures both effectiveness and efficiency.

The channel projection layer is implemented by a 1 × 1 convolution, which is essentially a fully connected operation and responsible for compressing and transforming information in the channel dimension.

The function of this layer is to enhance the expressiveness of features through nonlinear transformation, and reduce the overhead of subsequent calculation through channel compression. The use of LayerNorm ensures the stability of training, and the GELU activation function provides a smooth nonlinear transformation.

The spatial information mixture layer uses depth-wise separable convolution to capture local spatial context information.

Compared with the standard convolution, depth-wise separable convolution separates the spatial convolution from the channel convolution, which greatly reduces the number of parameters and computation while maintaining the same receptive field. The 3 × 3 convolution kernel size can effectively capture local spatial patterns, which is crucial for maintaining the continuity of the boundary.

The channel information mixing layer again uses a 1 × 1 convolution to perform the feature transformation of the channel dimension.

Finally, the input is added to the output via the residual join.

The residual connection not only alleviates the vanishing gradient problem, but also enables the network to focus on learning the residual mapping between input and output, which is particularly important in deep networks. Compared to a traditional convolutional decoder (e.g., the decoder in TransUNet with similar capacity), our SC-MLP decoder reduces parameters by approximately 42% (from 61.8 M to 35.8 M) and improves inference speed by 1.7× (from 26.3 fps to 44.8 fps) under the same input resolution and hardware, due to the efficient use of depthwise separable convolutions and pixel shuffling.

Optimization objective and loss function

Medical image segmentation, especially cardiovascular image segmentation, faces a serious problem of class imbalance. The blood vessel region usually only accounts for a very small proportion of the whole image, and this extreme imbalance make the traditional cross-entropy loss difficult to effectively optimize. Design a combinatorial optimization objective combining multiple loss functions to constrain the learning process of the model from different perspectives.

The overall loss function consists of three parts.

Where, λ = 0.6 and β = 0.2 are hyperparameters to balance the weights of different loss terms, these values are determined by grid search, which achieves the best balance across all metrics.

The Dice loss deals specifically with class imbalance and is defined as:

Where, p_i is the probability predicted by the model, g_i is the true label, N is the total number of pixels, and ε = 10^ -5 is the smoothing factor used to avoid division by zero errors. The Dice coefficient essentially measures the overlap between the predicted result and the true label. It is not sensitive to the number of foreground pixels, so it is particularly suitable for medical image segmentation tasks.

The cross-entropy loss provides pixel-wise supervision signals.

Where, C = 2 represents the number of classes (foreground and background), g_{i, c} is the true label of one-hot encoding, and p_{i, c} is the class probability predicted by the model. This loss term ensures that each pixel is correctly classified, providing fine-grained supervision information to the model. While cross-entropy loss alone has limited effectiveness when dealing with class imbalance, it can provide a beneficial complement when combined with other losses.

Margin loss specifically addresses the issue of boundary accuracy in cardiovascular segmentation tasks.

Here, ∇ denotes the gradient operator, which is used to calculate the spatial gradient of the predicted result and the true label. By calculating the difference between the predicted result and the true label in the gradient domain, the boundary loss forces the model to pay attention to the segmentation quality of the boundary region. In cardiovascular images, the accuracy of the boundary is crucial for subsequent quantitative analysis, such as stenosis rate calculation.

Results and discussion

Dataset and preprocessing

We used three publicly available cardiovascular datasets, such as ASOCA: 40 coronary CT angiography (CTA) scans with centerline and vessel wall annotations. Mean age of subjects is 67.5 ± 11.3 years (https://asoca.grand-challenge.org/). Cardiac-MRI (ACDC): 100 cardiac MRI scans from 100 patients, focusing on left ventricle segmentation. Mean age is 68.2 ± 12.1 years (https://www.creatis.insa-lyon.fr/Challenge/acdc/). Sunnybrook Cardiac Data: 45 cardiac MRI scans from 45 patients (age range 18–90 years, mean 62.1 ± 14.5), used as an additional external validation cohort to assess generalizability (https://www.cardiacatlas.org/sunnybrook-cardiac-data/.).

Preprocessing: All images were resampled to isotropic 1 × 1 × 1 mm³ resolution. Intensity normalization used z-score per case. A simple threshold-based connected component algorithm automatically localized the heart region to focus on the ROI. Data augmentation during training included random rotation (± 15°), random scaling (0.85–1.15), elastic deformation, and brightness/contrast adjustment.

Evaluation metrics and experimental setup

We used five metrics: Dice Similarity Coefficient (Dice), 95th percentile Hausdorff Distance (HD95), Average Symmetric Surface Distance (ASSD), Volume Overlap Error (VOE), Precision, and Recall. Additionally, for the ASOCA dataset, we report Centerline Dice¹ to specifically assess vessel topology preservation (Centerline Dice computes the Dice similarity coefficient on the vessel centerlines extracted from segmentation masks, measuring topological continuity).

Experimental Setup: Experiments were performed on 4×NVIDIA RTX 3090 GPUs using PyTorch 1.10. We used a 5-fold cross-validation strategy. For each fold, 80% of data was used for training, 10% for validation, and 10% for testing. Results are reported as mean ± standard deviation over the test folds. The AdamW optimizer was used with initial learning rate 1e-4, weight decay 1e-4, and a cosine annealing schedule with warm restarts (T_max = 300). Batch size was 16, trained for 300 epochs with early stopping (patience = 30). Mixed precision training was used for efficiency. Inference speed was measured on a single RTX 3090 with input size 256 × 256 and batch size 1, averaging over 1000 runs after 100 warm-up runs.

Table 1.

Configuration of experimental environment.

Categories	Configuration Items	Specific parameters
Hardware	GPU	4×NVIDIA RTX 3090(24GB)
	CPU	AMD Ryzen 9 5950X
	Memory	128GB DDR4
Software	Deep learning frameworks	PyTorch 1.10.0
	CUDA version	11.3
	Major dependency libraries	NumPy, OpenCV, SimpleITK

For training parameters setting, we use AdamW optimizer with initial learning rate set to 1e-4 and cosine annealing strategy with hot restart. The batch size is set to 16, and the total number of training rounds is 300. To prevent overfitting, we implemented an early stopping strategy and monitored model performance on the validation set. Figure 4 shows the loss function and evaluation metrics during training.

Fig. 4.

Loss and Dice coefficient during training, showing the convergence of the model after 200 rounds.

Comparison of experimental findings

We compared CTM-Net with six state-of-the-art methods, including UNet3 + and SegFormer-B3 (adapted for segmentation), and a ConvNeXt-Tiny U-Net. All comparison methods used official implementations and were trained under the same experimental conditions.

Table 2.

Performance comparison on ASOCA dataset. Best results in bold. Statistical significance vs. nnUNet is indicated (paired t-test, *p < 0.05).

Methods	Dice (%)	HD95 (mm)	ASSD (mm)		VOE (%)		Precision (%)		Recall (%)		Centerline Dice (%)		Params (M)		FLOPs (G)		Speed (fps)
U-Net	78.3 ± 1.2	12.5 ± 0.8	2.1 ± 0.3		21.7 ± 1.1		79.5 ± 1.3		77.8 ± 1.5		75.1 ± 1.8		34.9		62.3		45.2
Attention U-Net	79.1 ± 1.1	11.8 ± 0.7	1.9 ± 0.2		20.9 ± 1.0		80.2 ± 1.2		78.5 ± 1.4		76.2 ± 1.7		36.2		68.1		42.8
U-Net++	79.5 ± 1.0	11.5 ± 0.6	1.8 ± 0.2		20.5 ± 0.9		80.8 ± 1.1		78.9 ± 1.3		76.9 ± 1.6		39.1		75.4		40.1
TransUNet	80.2 ± 0.9	10.9 ± 0.5	1.7 ± 0.2		19.8 ± 0.8		81.3 ± 1.0		79.6 ± 1.2		77.8 ± 1.5		105.2		152.8		25.3
Swin-Unet	80.8 ± 0.8	10.5 ± 0.4	1.6 ± 0.1		19.2 ± 0.7		82.1 ± 0.9		80.2 ± 1.1		78.5 ± 1.4		62.8		98.2		32.1
UNet3+	80.5 ± 0.8	10.7 ± 0.5	1.6 ± 0.1		19.5 ± 0.7		81.8 ± 0.9		80.0 ± 1.1		78.1 ± 1.4		42.5		80.1		38.5
SegFormer-B3	81.2 ± 0.7	10.1 ± 0.4	1.5 ± 0.1		18.8 ± 0.6		82.5 ± 0.8		80.8 ± 1.0		79.2 ± 1.3		47.3		89.5		35.2
ConvNeXt-T U-Net	81.4 ± 0.7	9.9 ± 0.3	1.4 ± 0.1		18.6 ± 0.6		82.6 ± 0.8		81.1 ± 1.0		79.5 ± 1.3		52.1		95.3		33.8
nnUNet	81.5 ± 0.7	9.8 ± 0.3	1.5 ± 0.1		18.5 ± 0.6		82.7 ± 0.8		81.0 ± 1.0		79.8 ± 1.2		45.3		87.2		36.5
CTM-Net	82.1 ± 0.6*	9.3 ± 0.2*	1.3 ± 0.1*		17.9 ± 0.5*		83.5 ± 0.7		81.8 ± 0.9		80.9 ± 1.0*		35.8		58.5		44.8

CTM-Net achieves the best performance on all metrics. The Dice improvement over nnUNet (0.6%) is statistically significant (p = 0.023). CTM-Net also achieves superior boundary accuracy (HD95 9.3 mm) and topology preservation (Centerline Dice 80.9%). Notably, CTM-Net achieves this with significantly lower computational cost: 58.5G FLOPs compared to 87.2G for nnUNet and 152.8G for TransUNet, while also achieving the fastest inference speed (44.8 fps). Figure 5 shows qualitative comparisons, illustrating CTM-Net’s superior boundary accuracy and detail preservation.

Fig. 5.

Comparison of segmentation results of different methods on typical cases, showing the advantage of CTM-Net in boundary accuracy and detail preservation.

Figure 5 shows the visual comparison of the segmentation results of each method on typical cases. From the visual results, it can be seen that CTM-Net has better segmentation results at the blood vessel boundary.

Analysis of ablation experiments

In order to verify the contribution of each component in CTM-Net, we designed a systematic ablation experiment. Firstly, we constructed the baseline model, and then gradually added each component to observe the performance change. Table 3 shows the detailed ablation experiment results.

Table 3.

Results of ablation experiments for core components.

Experimental setup	Dice (%)	HD95 (mm)	Params (M)	FLOPs (G)	Speed (fps)
Baseline (CNN only)	78.3 ± 1.2	12.5 ± 0.8	35.2	62.1	45.2
+ Transformer module	80.2 ± 0.9	10.9 ± 0.5	38.7	68.3	38.5
+ MLP decoder	79.6 ± 1.0	11.2 ± 0.6	32.1	52.4	52.3
Complete CTM-Net	82.1 ± 0.6	9.3 ± 0.2	35.8	58.5	44.8

The Transformer module adds 1.9% Dice by improving global context understanding. The MLP decoder alone (with CNN encoder) maintains performance while reducing parameters and increasing speed. The complete model achieves the best performance, demonstrating synergistic benefits.

Table 4 shows that 6-layer transformers with 8 attention heads and learnable position encoders achieve the best performance balance. Too many or too few layers will affect the performance of the model, indicating that the appropriate network depth should be selected according to the complexity of the task. Figure 6 illustrates the convergence of different Transformer configurations during training. It can be seen that the 6-layer Transformer performs the best in terms of both convergence speed and final performance.

Table 4.

Transformer configuration ablation experimental results.

Configuration	Dice (%)	HD95 (mm)	Params (M)	Training time (h)
Layers
4-layer	81.3 ± 0.7	9.8 ± 0.3	36.2	28.5
6-layer	82.1 ± 0.6	9.3 ± 0.2	38.7	32.1
8-layer	81.8 ± 0.6	9.5 ± 0.3	41.3	36.8
Attention heads
4 heads	81.5 ± 0.7	9.6 ± 0.3	37.1	30.2
8 heads	82.1 ± 0.6	9.3 ± 0.2	38.7	32.1
12 heads	81.9 ± 0.6	9.4 ± 0.2	40.4	35.7
Positional encoding
Sinusoidal	81.7 ± 0.7	9.5 ± 0.3	38.7	31.8
Learnable	82.1 ± 0.6	9.3 ± 0.2	38.7	32.1

Fig. 6.

Variation curves of training loss and validation set dice coefficient for different transformer configurations.

In terms of MLP decoder design, we compare different upsampling methods and feature fusion strategies (Table 5).

Table 5.

Ablation study for MLP decoder design. “w/o spatial mixing” removes the depthwise convolution layer from the SC-MLP block, relying only on channel mixing.

Design choices	Dice (%)	HD95 (mm)	Params (M)	Speed (fps)
Upsampling method
Transposed Conv	81.3 ± 0.7	9.8 ± 0.3	38.2	40.2
Bilinear interp	81.5 ± 0.7	9.6 ± 0.3	35.1	48.5
Pixel shuffling	82.1 ± 0.6	9.3 ± 0.2	35.8	44.8
Fusion Strategy
Feature addition	81.7 ± 0.6	9.5 ± 0.3	35.3	46.2
Feature concat	82.1 ± 0.6	9.3 ± 0.2	35.8	44.8
SC-MLP components
w/o residual	81.2 ± 0.8	9.9 ± 0.4	35.1	45.3
w/o spatial mixing	81.5 ± 0.7	9.6 ± 0.3	34.8	47.1
Full SC-MLP	82.1 ± 0.6	9.3 ± 0.2	35.8	44.8

As seen in Table 5, it can be found that the pixel shuffle upsampling method outperforms the traditional method in performance while maintaining a high inference speed. The feature concatenation strategy achieves better performance than the feature addition strategy, which indicates that retaining more feature information is crucial to the segmentation accuracy. The complete SC-MLP block design works best when the individual components work together. Figure 7 shows the comparison results of different design choices in boundary segmentation quality, showing that the complete design of CTM-Net is more accurate and smooth in vessel boundary segmentation.

Fig. 7.

The visual comparison of different MLP decoder design choices on vessel boundary segmentation.

Through systematic ablation experiments, we verified the necessity and effectiveness of each component in CTM-Net. The experimental results shows that the co-design of CNN-Transformer-MLP can give full play to their respective advantages and achieve excellent performance while maintaining high efficiency.

Discussion

CTM-Net demonstrates superior performance and efficiency for cardiovascular image segmentation. The synergistic design of CNN, Transformer, and MLP effectively balances local detail, global context, and computational cost. The lightweight SC-MLP decoder is a key enabler of this efficiency. Statistical validation confirms the significance of improvements over strong baselines. However, we acknowledge limitations. The datasets, while diverse, are relatively small and do not exclusively comprise elderly patients. The performance gains, while significant, are incremental. Most importantly, clinical applicability requires further validation, including external validation on large, multi-center cohorts. Correlation of segmentation results with clinically relevant parameters (e.g., stenosis quantification accuracy, plaque burden estimation). Integration with clinical workflows and assessment by clinicians and prospective studies to evaluate impact on diagnosis and treatment planning.

Conclusion

This paper presented CTM-Net, a novel CNN-Transformer-MLP collaborative framework for cardiovascular image segmentation. The method achieves state-of-the-art performance on public datasets with high computational efficiency. The synergistic design and lightweight MLP decoder offer a promising technical pathway. Future work will focus on prospective clinical validation, integration with stenosis quantification pipelines, and extension to other cardiovascular imaging modalities such as intravascular ultrasound and optical coherence tomography.

Author contributions

All authors contributed substantially to the conception and design of this article. Bian Chen collected and analyzed the data, and wrote the article. Li Wei, Peng Longhua, Zhao Cheng, Cheng Cheng and Ma Fangfang critically revised the manuscript and provided the facilities. Li Wei, Peng Longhua, Zhao Cheng, Cheng Cheng and Ma Fangfang analyzed the data.

Funding

The research fund is supported by the Xuzhou Science and Technology Project 2023 Medical and Health General Project, China (KC23168).

Data availability

The datasets used and/or analysed during the current study available from the corresponding author on reasonable request.

Declarations

Competing interests

The authors declare no competing interests.