{S$^3$-Mamba}: Small-Size-Sensitive Mamba for Lesion Segmentation

Abstract: Small lesions play a critical role in early disease diagnosis and intervention of severe infections. Popular models often face challenges in segmenting small lesions, as it occupies only a minor portion of an image, while down_sampling operations may inevitably lose focus on local features of small lesions. To tackle the challenges, we propose a {\bf S}mall-{\bf S}ize-{\bf S}ensitive {\bf Mamba} ({\bf S$^3$-Mamba}), which promotes the sensitivity to small lesions across three dimensions: channel, spatial, and training strategy. Specifically, an Enhanced Visual State Space block is designed to focus on small lesions through multiple residual connections to preserve local features, and selectively amplify important details while suppressing irrelevant ones through channel-wise attention. A Tensor-based Cross-feature Multi-scale Attention is designed to integrate input image features and intermediate-layer features with edge features and exploit the attentive support of features across multiple scales, thereby retaining spatial details of small lesions at various granularities. Finally, we introduce a novel regularized curriculum learning to automatically assess lesion size and sample difficulty, and gradually focus from easy samples to hard ones like small lesions. Extensive experiments on three medical image segmentation datasets show the superiority of our S$^3$-Mamba, especially in segmenting small lesions. Our code is available at https://github.com/ErinWang2023/S3-Mamba.

• The paper introduces S3-Mamba, advancing small lesion segmentation with an Enhanced Visual State Space block and tensor-based cross-feature multi-scale attention to preserve key image details.
• It employs a regularized curriculum learning strategy that gradually refines training by prioritizing challenging, small lesion samples.
• Experimental results on ISIC2018, CVC-ClinicDB, and Lymph datasets demonstrate superior mIoU and accuracy compared to traditional models, promising better early diagnosis.

Overview of S$^3$-Mamba: Small-Size-Sensitive Mamba for Lesion Segmentation

The paper introduces S$^3$-Mamba, a novel approach focused on enhancing the accuracy of small lesion segmentation in medical imaging. This model aims to overcome the prevalent challenges that existing segmentation techniques face, particularly promoting sensitivity to small lesions across channel, spatial, and training dimensions. The research emphasizes the crucial importance of accurately identifying small lesions, which are often the earliest indicators of severe diseases and critical for early diagnosis and intervention.

Key Contributions and Methodology

1. Enhanced Visual State Space Block (EnVSSBlock):
   • The S$^3$-Mamba introduces an Enhanced Visual State Space block that leverages residual connections and channel-wise attention to preserve critical local features while selectively amplifying important lesion details. This approach allows the model to retain crucial details often lost during traditional down-sampling processes, particularly affecting small lesion segmentation.
2. Tensor-based Cross-feature Multi-scale Attention (TCMA):
   • This component exploits multi-scale feature integration, combining image input, intermediate layer, and edge features to maintain spatial details at varied granularities. The TCMA facilitates cross-feature attention over different scales, thereby enhancing the model's ability to discern and segment small lesions accurately across diverse medical imaging datasets.
3. Regularized Curriculum Learning Strategy:
   • The paper introduces a novel training strategy that dynamically assesses lesion size and sample difficulty, gradually transitioning the focus from simpler to more complex samples. This curriculum learning approach is regulated to prioritize challenging samples, such as small lesions, enhancing the training efficiency and segmentation precision.

Experimental Evaluation and Results

The S$^3$-Mamba demonstrates superior performance across multiple medical imaging datasets, notably in segmenting small lesions. It outperforms traditional models like UNet, and newer transformer and state-space models across key metrics: mean Intersection over Union (mIoU), Dice Similarity Coefficient (DSC), Accuracy (ACC), Specificity (SPE), and Sensitivity (SEN).

• In the ISIC2018 dataset, the S$^3$-Mamba achieved notable improvements in the segmentation of small lesions with a mIoU of 77.13%, highlighting its efficacy in scenarios where lesion visibility might be low due to size constraints.
• The CVC-ClinicDB dataset results further underscore its prowess, achieving an mIoU of 75.40% for small lesions and maintaining high accuracy across larger lesion categories.
• The model also excels in ultra-small lesion segmentation, as evidenced in the private Lymph dataset results, clearly demonstrating its robustness and adaptability to diverse medical imaging conditions.

Implications and Future Directions

Practically, the S$^3$-Mamba's ability to accurately segment small lesions holds significant potential for improving early diagnosis and treatment planning, potentially reducing the onset severity of various diseases. Theoretically, this work adds valuable insights to the ongoing development of state-space models and attention mechanisms tailored for biomedical image analysis, paving the way for further exploration of hybrid models that combine spatial, channel, and training strategies.

Future research may benefit from exploring broader modalities of data, further optimizing computational efficiency, and expanding the application of S$^3$-Mamba to other forms of small object segmentation in different contexts of medical imaging. Additionally, investigating the integration of this model with diagnostic workflows could yield practical benefits for healthcare professionals by streamlining and enhancing diagnostic accuracy.