MALUNet: A Multi-Attention and Light-weight UNet for Skin Lesion Segmentation (2211.01784v1)

Abstract: Recently, some pioneering works have preferred applying more complex modules to improve segmentation performances. However, it is not friendly for actual clinical environments due to limited computing resources. To address this challenge, we propose a light-weight model to achieve competitive performances for skin lesion segmentation at the lowest cost of parameters and computational complexity so far. Briefly, we propose four modules: (1) DGA consists of dilated convolution and gated attention mechanisms to extract global and local feature information; (2) IEA, which is based on external attention to characterize the overall datasets and enhance the connection between samples; (3) CAB is composed of 1D convolution and fully connected layers to perform a global and local fusion of multi-stage features to generate attention maps at channel axis; (4) SAB, which operates on multi-stage features by a shared 2D convolution to generate attention maps at spatial axis. We combine four modules with our U-shape architecture and obtain a light-weight medical image segmentation model dubbed as MALUNet. Compared with UNet, our model improves the mIoU and DSC metrics by 2.39% and 1.49%, respectively, with a 44x and 166x reduction in the number of parameters and computational complexity. In addition, we conduct comparison experiments on two skin lesion segmentation datasets (ISIC2017 and ISIC2018). Experimental results show that our model achieves state-of-the-art in balancing the number of parameters, computational complexity and segmentation performances. Code is available at https://github.com/JCruan519/MALUNet.

• The paper introduces MALUNet, a novel multi-attention model that reduces computational complexity while significantly improving mIoU and DSC scores.
• It integrates key modules like Dilated Gated Attention and Inverted External Attention to effectively capture and fuse multi-scale features.
• Experimental results on ISIC2017/2018 datasets show MALUNet outperforms traditional UNet with a 2.39% mIoU and 1.49% DSC improvement using 44x fewer parameters.

MALUNet: Advancements in Skin Lesion Segmentation

The academic paper titled MALUNet: A Multi-Attention and Light-weight UNet for Skin Lesion Segmentation presents a novel approach that addresses computational challenges in clinical environments through a light-weight segmentation model. The authors propose MALUNet, a model designed to enhance performance metrics such as mean Intersection over Union (mIoU) and Dice similarity score (DSC) while drastically reducing the number of parameters and computational complexity compared with traditional UNet architectures.

Proposed Modules and Model Architecture

The paper introduces four key modules integrated into a U-shape architecture to facilitate efficient medical image segmentation:

1. Dilated Gated Attention (DGA): This module extracts local and global feature information using dilated convolutions combined with gated attention mechanisms. By applying depthwise separable convolutions with varied dilation rates, DGA enables reduced parameters while maintaining competitive segmentation precision.
2. Inverted External Attention (IEA): Enhancing intra-dataset feature associations and characterizations, IEA employs two memory units with broadened expansion channels compared to traditional external attention setups, leveraging a higher dimensional space to depict comprehensive dataset characteristics.
3. Channel Attention Bridge Block (CAB): Focusing on channel-level attention maps, CAB fuses features from multiple stages using global average pooling and 1D convolutions, followed by fully connected layers. This module enables effective integration of multi-stage feature information.
4. Spatial Attention Bridge Block (SAB): SAB implements pooling operations in the channel dimension to generate spatial attention maps, using shared dilated convolutions. It focuses on fusing multi-scale information across spatial dimensions to produce attentive feature outputs.

These modules integrate within the MALUNet architecture, resulting in a model with reduced computational demands. The six-stage U-shape architecture configures channel sizes at {8, 16, 24, 32, 48, 64}, and combines attention mechanisms strategically to develop a robust segmentation model. The sequential arrangement of SAB followed by CAB ensures effective multi-scale and cross-stage feature fusion.

Experimental Results

Evaluations performed on ISIC2017 and ISIC2018 datasets demonstrate the efficacy of MALUNet, achieving noticeable improvements in mIoU and DSC scores. Despite its reduced parameter count and complexity, MALUNet achieves superior segmentation performance:

• ISIC2017 Dataset: MALUNet improved mIoU by 2.39\% and DSC by 1.49\% over UNet, with a significant reduction in parameters (44x) and computational complexity (166x).
• ISIC2018 Dataset: MALUNet matches the competitive performance of larger models like TransFuse, showcasing its capability to achieve satisfactory results in real-world datasets, while maintaining its lightweight nature.

Implications and Future Directions

The results suggest significant implications for mobile health applications and computational efficiency in clinical environments. MALUNet's architecture could be expanded to address other medical imaging tasks beyond skin lesions. Future research could explore further optimization through techniques like model pruning and neural architecture search to enhance both versatility and applicability in broader medical domains.

Theoretical advancements lie in dynamically assessing attention mechanisms to support generalizable network designs, potentially introducing transformer-based components more broadly across medical image analysis tasks. With these foundations, the principles underlying MALUNet could inspire innovative pathways for developing high-efficiency AI models suited to diverse clinical scenarios.