Brain tumor is a mass formed by the uncontrolled proliferation of brain cells, which occurs in males and females of all ages and is one of the most dangerous diseases in the world [1]. Therefore, early identification of brain tumors is particularly crucial to improve the treatment effect and survival rate of patients. Among various imaging techniques, magnetic resonance imaging (MRI) is currently the most popular method for detecting brain tumors [2]. From the perspective of MRI, brain tumors can be further classified into gliomas, meningiomas, and pituitary tumors based on their size, shape, and location [3]. Each type of tumor can be life-threatening. Therefore, it is crucial to classify different types of brain tumors effectively and then conduct targeted treatments for each type. However, due to the small structural differences between different brain tumors, accurately classifying them is challenging. Manual classification by doctors is inevitably prone to misdiagnosis and inefficient, while also increasing the burden on doctors. To overcome these difficulties, machine learning-based techniques have begun to be applied to MRI images for automatic brain tumor classification tasks and have played a crucial role in computer-aided diagnosis (CAD) systems.

So far, the emergence of CNN has revolutionized the field of medical image processing. Traditional machine learning algorithms such as Support Vector Machine (SVM) and Decision Tree often require manual feature design, which relies on domain knowledge and expert experience, and the process is cumbersome. In contrast, CNN can automatically learn feature representations from raw data, reducing the reliance on manual feature engineering. Therefore, CNN have significantly improved the performance of CAD systems. Although CNN have achieved better classification performance compared to traditional machine learning methods, due to the characteristics of local feature extraction, CNNS may ignore some key inherent tumor properties when processing brain tumor classification tasks, such as the contextual information surrounding the tumor region and the size variations of the tumor. The lack of these pieces of information can lead to limitations in the model’s ability to identify tumors. Additionally, the high similarity between brain tumor categories add to additional difficulties for the practical application of CNN. To overcome these challenges, researchers [4–6] have increasingly integrated attention mechanisms into CNN models, By incorporating attention mechanisms, CNNs are able to capture finer-grained details within tumor regions, which are crucial for distinguishing between different types of brain tumors.

The emergence of vision transformer (ViT) [7] has overcome the limitations of CNN, yet it require extensive computational resources during training, limiting the input sequence length and increasing training time. Recently, Mamba [8–10] has not only overcome these difficulties but also possesses the ability to extract global features. The Spatial State Model (SSM) represented by Mamba operates in linear time with respect to sequence length. Benefiting from its linear time complexity, Mamba has a lower computational cost, potentially demonstrating higher computational efficiency in handling complex tasks.

Despite significant advancements in deep learning for medical image classification, existing studies still fall short in capturing subtle distinctions in brain tumors and reducing computational costs. Traditional CNN perform well in local feature extraction but are limited in handling long-range dependencies and global features. Additionally, while ViT overcome the limitations of CNN, they require substantial computational resources. The main objective of this study is to develop an efficient brain tumor classification model that achieves high classification accuracy, with the aim of enhancing the practical applicability of automatic brain tumor detection and diagnosis. To this end, we propose a novel parallel architecture named ParMamba, which utilizes ConvMamba blocks as its backbone and incorporates Convolutional Attention Patch Embedding (CAPE). The ConvMamba block combines the local feature extraction capabilities of CNN with the global feature capturing strengths of the Mamba, resulting in improved computational efficiency and classification accuracy. Furthermore, the channel enhancement module is introduced to enhance cross-channel feature interactions. CAPE captures brain tumor information from multiple perspectives by applying both max pooling and convolutional pooling in two directions, which enriches the feature representation. The main contributions of this paper are summarized as follows:

This paper propose a parallel architecture called ParMamba, designed for effective multi-class classification of brain tumors in MRI images.

Brain tumors exhibit a variety of types, each with distinct characteristics and manifestations. Furthermore, many brain tumors have indistinct boundaries, making them easily confusable with surrounding normal tissue, which adds to the complexity of classification. In early automated systems, people used machine learning methods to identify tumors of brain tumors from MRI images. El-Dahshan et al. [11] used discrete wavelet transform (DWT) to obtain features related to MRI images and classified them using a k-nearest neighbor-based classifier. Shim et al. [12] combined finite element analysis and machine learning methods for detecting brain injuries. Das et al. [13] adopted Ripplet Transform Type-I (RT) to represent significant features of brain MRI images and used least squares support vector machines to classify brain MRI images. Zhang et al. [14] compared traditional training methods such as Scaled Chaotic Artificial Bee Colony (SCABC), momentum BP, genetic algorithms, and simulated annealing, indicating that the SCABC method is better. Shim et al. [15] developed an efficient computational pipeline to generate finite element models of brain injury for specific individuals, predicting structural damage following trauma. However, traditional machine learning still poses challenges due to its reliance on manual intervention and the requirement for extensive domain knowledge and expertise.

CNN can automatically learn feature representations from raw data, reducing the reliance on manual feature engineering. Therefore, CNN has been widely used in CAD methods for medical images and has revolutionized the field of medical image analysis. Ayadi et al. [16] proposed a deep CNN model, utilizing ten different convolutional layers to extract features from brain tumor images, aiming to enhance the ability to capture brain tumor features. Atha et al. [17] employed CNN as the base architecture and incorporated the idea of semi-supervised learning, combining the training process of labeled and unlabeled brain tumor data, enabling the network to learn from both types of data simultaneously. Rizwan et al. [18] used Gaussian convolution kernels to extract local features of brain tumors, enhancing the accuracy of brain tumor feature extraction. Zhu et al. [19] adopted DenseNets as the base network structure and introduced a biologically inspired evolutionary algorithm to optimize the network structure of DenseNets, adapting it to the characteristics of medical image data. Aamir et al. [20] used two pre-trained CNN networks, EfficientNet and ResNet50, to extract features from brain tumor images separately. Then, they employed partial least squares (PLS) to fuse the feature vectors extracted from the two models, forming a hybrid feature vector. Kumar et al. [21] proposed a multi-class brain tumor classification method based on ResNet-50 and global average pooling, which performs well in addressing gradient disappearance and overfitting issues in deep networks. Gursoy et al. [22] integrated graph neural networks (GNNs) to capture relational dependencies between image regions and CNN to extract spatial features, enhancing the detection of brain tumors.

However, when using CNN models for brain tumor classification, limited data can lead to the over-fitting problem. Therefore, to address the issue of small brain tumor image samples, data augmentation techniques have been applied to enlarge the dataset in some works. Yaqub et al. [23] utilized generative adversarial networks (GANs) for data augmentation. Ghassemi et al. [24] employed GANs for preprocessing brain tumor images, in addition to data augmentation methods such as image rotation and mirroring. Li et al. [25] adopted data augmentation techniques including rotation, horizontal flipping, and vertical flipping, as well as salt-and-pepper noise as a data augmentation method.

In recent years, with the introduction of attention mechanism into the field of computer vision, researchers have also combined attention mechanism with CNN and applied it to medical images. Dutta et al. [26] adopted CNN as the base architecture and introduced a lightweight global attention mechanism after the CNN layers, enhancing the model’s ability to extract more salient features from brain tumor images. Wang et al. [27] utilized a pre-trained vision transformer as the base architecture, and combined with token merging techniques, to extract key information from brain tumor images. Isunuri et al. [28] employed a pre-trained efficientNetB4 to extract brain tumor features and then used multi-path convolution and multi-head attention for feature enhancement.

More recently, Mamba based on state-space models (SSM) has emerged in people’s vision. Mamba not only has the ability of global feature extraction, but also exhibits linear complexity related to the size of the input image. Yue et al. [29] combined the local feature extraction capability of CNN with the ability of SSM to capture long-range dependencies, thereby modeling medical images in different modes. Ma et al. [30] have utilized the integration of CNN’s local pattern recognition capabilities with Mamba’s global context understanding, enabling automatic adaptation to various datasets and applicability in segmentation tasks across diverse biomedical imaging fields. Ruan et al. [31] have proposed a medical image segmentation model based on SSM. By leveraging Visual State Space (VSS) blocks to capture extensive contextual information, they have constructed an asymmetric encoder-decoder architecture, marking it as the first medical image segmentation model purely built upon SSM.

The architecture of the proposed ParMamba adopts four units for brain tumor classification, which is similar to numerous prior studies [29,32,33]. As shown in Fig. 1, Each unit incorporates a patch embedding layer, succeeded by multiple sequentially arranged ConvMamba blocks. Specifically, the patch embedding layer adopts convolutional attention patch embedding layer [32], while each ConvMamba block consists of a ConvBlock, a MambaBlock, and a channel enhancement module.

In the first unit, The channel dimension of the input image x∈R3×H×W is mapped to C1 channels via convolutional attention patch embedding layer, obtaining the embedded image x1′∈RC1×H4×W4. Then, the x1′ is fed into N1 sequentially stacked ConvMamba blocks to extract image features, yielding X1′∈RC1×H4×W4. The second, third, and fourth units repeat the above operations, resulting in the fourth unit, the X4′∈RC4×H32×W32 is obtained. Finally, classifier predicts whether the input image is glioma, meningioma, notumor, or pituitary.

This section provides a detailed description of the experiments conducted in this paper, including the experimental details, and the datasets used in the experiments. Comparisons are also made with existing brain tumor classification methods. Furthermore, ablation studies were performed to investigate the effectiveness of CAPE, MambaBlock, and the channel enhancement module.

This paper proposes a novel parallel architecture named ParMamba for brain tumor classification. The ParMamba consists of four components: ConvBlock, MambaBlock, channel enhancement module, and CAPE. Among them, ConvBlock and MambaBlock extract local and global features respectively, channel enhancement module implements cross-channel communication, and CAPE is used for downsampling, effectively extracting fine-grained features of brain tumors from local and global contexts. Extensive experiments are conducted on two publicly available brain tumor datasets. Dataset 1 contains four categories of images: glioma, meningioma, pituitary, and non-tumor, while dataset 2 comprises three categories: glioma, meningioma, and pituitary. The experimental results demonstrate that the proposed ParMamba achieves outstanding classification performance, with accuracies of 99.62% and 99.35% on dataset 1 and dataset 2, respectively. This indicates that the model is capable of achieving excellent classification performance on both datasets. Compared to existing brain tumor classification methods, ParMamba marks a significant advancement in the field. Moreover, the accuracy of Mamba increased by 1.37% compared with ViT, while FLOPs decreased by 55%, and throughput rate also increased by more than 30% compared with ViT. This indicates that the proposed ParMamba is not only faster but also more accurate than attention mechanisms. Furthermore, ablation studies are conducted to validate the effectiveness of CAPE, MambaBlock, and channel enhancement module individually. The results reveal that the combination of CAPE, MambaBlock, and channel enhancement module effectively extracts brain tumor features, leading to excellent classification results.

While the proposed ParMamba model demonstrates strong overall performance, its recognition of meningioma remains relatively limited. Therefore, future efforts could focus on refining the model to better capture meningioma features. Moreover, future research could extend the application scope of ParMamba by testing its generalization performance on large-scale brain tumor datasets. Further potential directions include developing lightweight network designs for efficient deployment on resource-constrained devices.