Tiny machine learning (TinyML) for healthcare: Biomedical imaging segmentation, which involves accurately delineating anatomical structures and pathological regions from medical images, is essential for precise diagnostics. Recent advances in AI have significantly improved segmentation methods, enhancing both their accuracy and efficiency. A burgeoning area of research in this field is the integration of TinyML into healthcare, particularly for applications like lesion segmentation, offering exciting possibilities for both research and practical use.

TinyML refers to deploying machine learning models on low-power, compact hardware. This technology holds the potential to revolutionize healthcare by enabling advanced analytical capabilities directly at the point of care. It facilitates real-time, on-device processing, making sophisticated medical image analysis feasible even in settings with limited traditional computing resources or in mobile healthcare environments. For instance, using TinyML for lesion segmentation could provide immediate diagnostic feedback during patient consultations or in remote locations, significantly reducing dependence on extensive infrastructure typically needed for detailed analyses.

The incorporation of TinyML into medical devices is expected to enhance diagnostic accuracy, improve patient outcomes, and broaden access to advanced medical technologies in underserved regions. To optimize the deployment of TinyML in these critical applications, researchers are exploring advanced methods such as hyper-structure optimization [22] and leveraging efficient techniques like binary neural networks [23].

Hyper-structure optimization focuses on reducing the model’s parameter count without sacrificing performance, ensuring the models remain both practical and lightweight for use on miniature devices. Moreover, implementing binary neural networks helps streamline computations, further enhancing the practicality of TinyML applications in resource-constrained settings. As we delve into optimizing and applying models like MUCM-Net in healthcare. This research not only underscores the transformative possibilities of TinyML but also guides future explorations in deploying compact, efficient AI solutions in medical settings.

Supervised methods of segmentation: As AI technology continues to advance, the approaches for medical image segmentation have evolved significantly. Initially, the field heavily relied on CNNs such as U-Net, which is essential in medical image segmentation [24], and its attention-enhanced variant, Att-UNet [25], which incorporates attention mechanisms to further refine the segmentation accuracy by focusing on relevant features within the images. The development of hybrid architectures marks a further evolution in segmentation techniques. There are some hybrid-based UNets for medical image segmentation: (1) Transformers-related: such as TransUNet [26], TransFuse [27], and SANet [28]; (2) MLP-related: such as ConvNeXts [29], UNeXt [30], MALUNet [31] and its extended version EGE-UNet [32]. Recently, as Vision Mamba [33]’s image processing ability with fewer parameters and lower computations, Mamba-based hybrid structure UNets are becoming popular such as VM-UNet [34], VM-UNet V2 [35], LightM-UNet [36], and UltraLight VM-UNet [37].

State-space models (SSMs): SSMs have recently been recognized for their linear complexity concerning input size and memory usage, establishing them as fundamental components for lightweight model architectures [38]. SSMs are particularly effective at capturing long-range dependencies, offering a critical solution to the convolution challenge of processing information across extensive distances. With the advantage of SSMs, Mamba [39] has been proven to handle textual data with fewer parameters than Transformers. Similarly, the advent of Vision Mamba has advanced the application of SSMs in image processing, demonstrating a significant memory reduction, all without relying on traditional attention mechanisms. This pioneering research bolsters confidence in Mamba’s potential as a critical lightweight model component in future technological advancements.

In this paper, we extend our previous work UCM-Net to propose a new hybrid work MUCM-Net which engages the Mamba’s features learning ability and maintains fewer parameters and lower computations.

Network structure design: Figure 3 presents a detailed overview of the structural framework of MUCM-Net, a cutting-edge architecture featuring a unique U-shape design, building upon UCM-Net. MUCM-Net consists of a down-sampling encoder and an up-sampling decoder, forming a robust network for skin lesion segmentation. The network is composed of six encoder-decoder stages, with channel capacities set at {8, 16, 24, 32, 48, 64}. The first stage in the encoder-decoder sequence is a convolutional block, responsible for extracting and capturing critical features. The subsequent stages incorporate our novel UCM-Net blocks, enhancing the network’s performance.

Convolution block: In our design, the first encoder-decoder stage employs a standard convolution layer with a 3 × 3 filter. This convolution block utilizes a 3 × 3 kernel, which is widely used to capture spatial relationships within input features. This kernel size is especially beneficial in the initial layers of the network, where maintaining the spatial integrity of feature maps is crucial for decoding complex input patterns. From the 2nd to the 6th stage, we incorporate a 1 × 1 filter convolution layer, which supports the subsequent Mamba-UCM block. This adjustment significantly reduces the number of learnable parameters and the overall computational load, optimizing the network’s efficiency.

Mamba-UCM block: The 2nd–6th stages mainly use the Mamba-UCM block for feature learning. The Mamba-UCM block showcases an advanced strategy that merges UCM-Net block, which contains CNNs with MLPs and Mamba block, to enhance feature learning. There are two advantages in the hybrid structure: (1) CNNs module can enable the model to extract the spatial feature; (2) MLPs and SSM can power the model’s pattern recognition capabilities. The process begins by reshaping the initial input feature map to meet the distinct requirements of CNNs, MLPs, and Mamba. This adaptation involves converting a four-dimensional tensor suitable for CNN processing into a three-dimensional tensor appropriate for MLP and Mamba operations. Inspired by UltraLight VM-UNet and Vision Mamba, we proposed four versions of MUCM-Net with different patch processing. Figure 4 presents the visible differences between the UCM-Net block and the two visions of Mamba-UCM blocks. The PyTorch-style pseudocode in Figure 5 presents our defined sequence of operations, which is how we combine the UCM block and the Mamba block for feature learning.

In our solution, we designed a new group loss function similar to those used in TransFuse, EGE-UNet, and our previous work, UCM-Net. However, different from theirs, our proposed base loss function is calculated from Binary Cross-Entropy (BCE) loss (1) and Dice loss (2) components and square Dice loss (3) components to calculate the loss from the scaled layer masks in different stages compared with the ground truth masks.

The advantages of using different loss functions are as follows:

1.

BCE loss: Widely used in classification tasks, including biomedical image segmentation, these losses are highly effective for pixel-level segmentation.

where N is the total number of pixels (for image segmentation) or elements (for other tasks), y_i is the ground truth value, and p_i is the predicted probability for the i-th element.

where we used a smooth constant, which equals 1e-5, to enhance numerical stability.

where the equation represents an improvement over the standard Dice loss by placing greater emphasis on the squared terms of intersections and unions.

Equations 1, 2, and 3 define the base loss function (Equation 4) for our proposed model, incorporating the Dice loss, and square Dice loss components. λ_i represents the weight assigned to different stages. We explore adding the difference between the up-sample from the stage with ground truth to the final loss. Because the large stage’s result should be close to the final prediction, we set λ_i to 0.1, 0.2, 0.3, 0.4, and 0.5 based on the i-th stage, as illustrated in Figure 3.

Equation 7 defines our proposed group loss function, which calculates the loss based on the scaled layer masks at various stages compared to the ground truth masks. Equations 5 and 6 detail the stage loss for different intermediate layers and the output loss for the final layer, respectively.

To prove the effectiveness of our proposed methods, we did an ablation experiment of various base loss functions and the experimental results are reported in the following Discussion section.

Tables 1 and 2 provide a thorough evaluation of the prediction performance of MUCM-Net, our novel Mamba-based skin lesion segmentation model, in comparison to established models. This assessment is conducted using the widely recognized ISIC2017 and ISIC2018 datasets. Introduced in 2024, MUCM-Net proves to be a robust and highly competitive solution in this domain. The key takeaway from these Tables is MUCM-Net’s ability to outperform all previous models, establishing a new state-of-the-art for lightweight skin lesion segmentation. Our model presents exceptional performance across various prediction metrics, underscoring its advancement in the field and potential to redefine the standard for accurate skin lesion delineation. To assess the physical performance of our proposed model, we compared its computational aspects and number of parameters with those of different segmentation models, as shown in Table 3. Remarkably, MUCM-Net (8-patch) operates with lower GFLOPS compared to other Mamba-based models. This efficiency does not come at the cost of performance, as MUCM-Net maintains high accuracy and robustness in segmentation tasks with the Mamba structure.

To prove the effects of our proposed loss function, we conducted an ablation experiment using different base loss functions to train our model, MUCM-Net (1-patch). Table 4 presents the experimental results on the ISIC2017 dataset, providing a comparative analysis of the performance metrics for each loss function. The base loss functions evaluated include BCE loss (A), Dice loss (B), and square Dice loss (C). Additionally, combinations of these loss functions were tested: A + B, A + C, B + C, and A + B + C. The performance metrics considered were the DSC, SE, SP, and ACC. The training settings and hardware environment are the same as the benchmarks in Tables 1 and 2.

Table 4 reveals that individual loss functions and their combinations exhibit varying performance on the ISIC2017 dataset. Among the individual loss functions, BCE loss (A) demonstrates a solid overall performance with a high SP of 0.9867 and ACC of 0.9680. However, its SE is relatively lower at 0.8878. Dice loss (B) improves SE to 0.9008 and maintains high ACC (0.9681), while square Dice loss (C) further enhances SE to 0.9051 and retains high SP (0.9831) and ACC (0.9684). Combinations of these loss functions generally yield better results, with A + B and A + C showing improvements in DSC and SE. The proposed combination of A + B + C achieves the best overall performance, with the highest DSC (0.9160), SE (0.9089), SP (0.9868), and ACC (0.9690). This indicates that integrating all three loss functions leverages their strengths, resulting in a more robust model performance across all evaluation metrics.

In conclusion, this paper introduces MUCM-Net, an innovative and efficient solution that integrates CNN, MLP, and Mamba to offer robust feature learning while keeping parameter counts low and computational demands minimal. Applied to the challenging task of skin lesion segmentation, MUCM-Net has been rigorously tested across various evaluation metrics, demonstrating superior performance compared to other recent lightweight or Mamba-based models.

Looking ahead, we plan to extend the application of MUCM-Net to other critical medical imaging tasks, aiming to advance the field and explore its potential across a broader range of healthcare applications. Additionally, we plan to explore how MUCM-Net can be effectively integrated with established hand-crafted segmentation methods (e.g., from [45, 46]) to leverage their complementary strengths and potentially achieve even higher segmentation accuracy. We will also investigate methods to address adversarial noise attacks on skin cancer segmentation models [47, 48], enhancing MUCM-Net’s robustness against potential manipulations that could compromise its performance. Our goal is to push the boundaries of deep learning in healthcare applications, making significant contributions to medical imaging as well as other areas of healthcare.