MALUNet: A Multi-Attention and Light-weight UNet for Skin Lesion Segmentation

Jiacheng Ruan, Suncheng Xiang, Mingye Xie, Ting Liu, Yuzhuo Fu

Introduction

In the field of medical imaging, the accurate segmentation of skin lesions is a critical precursor to effective diagnosis and treatment. While recent advancements in deep learning—particularly those utilizing Transformer-based architectures—have achieved remarkable accuracy, they often come at a significant computational cost. These "heavy" models require substantial memory and processing power, rendering them impractical for deployment in typical clinical settings or on mobile medical hardware.

To bridge this gap, the researchers behind MALUNet propose a highly efficient, lightweight neural network specifically designed for skin lesion segmentation. MALUNet aims to deliver state-of-the-art performance while maintaining a fraction of the computational footprint of traditional models, making it an ideal candidate for edge-device deployment.

Addressing the Challenges of Skin Lesion Segmentation

Skin lesion segmentation presents three primary technical hurdles that standard architectures often struggle to overcome:

  1. Local vs. Global Context: Models must balance the need for fine-grained local detail (to maintain sharp lesion edges) with a broad global context (to understand the lesion's overall structure).
  2. Inter-sample Relationships: Most segmentation models process images in isolation, failing to leverage shared characteristics and patterns present across the wider dataset.
  3. Multi-scale Targets: Lesions vary significantly in size and shape, necessitating a mechanism that can effectively fuse information across different stages of the network.

The MALUNet Architecture

MALUNet is built upon a 6-stage U-shaped architecture. Its efficiency is derived from the replacement of standard, computationally expensive convolutions with four novel, lightweight attention modules:

Dilated Gated Attention (DGA)

The DGA module utilizes depthwise separable convolutions with varying dilation rates. By spreading out the pixels the filter examines, the model can capture both fine local details and broad global context simultaneously. This approach enhances the receptive field without a corresponding explosion in parameter count.

Inverted External Attention (IEA)

Inspired by the principles of MobileNetV2, the IEA module incorporates two external "memory units" that learn features across the entire dataset. Unlike standard self-attention, which only considers the current image, IEA allows the model to "remember" common lesion characteristics. It employs an "inverted" process—expanding feature dimensions to capture richer information before compressing them back down—to maintain efficiency.

Channel and Spatial Attention Bridges (CAB & SAB)

In a standard UNet, the "skip connections" between the encoder and decoder can suffer from a semantic gap. MALUNet introduces the Channel Attention Bridge (CAB) and Spatial Attention Bridge (SAB) to act as sophisticated intermediaries. These modules fuse features from multiple stages of the network before they reach the decoder, ensuring that the model handles multi-scale targets with high precision.

Experimental Results

The model was rigorously evaluated on two gold-standard public datasets: ISIC2017 (2,150 images) and ISIC2018 (2,694 images). All images were resized to 256x256 for consistency, using a 7:3 training-to-testing split.

Performance vs. Efficiency

MALUNet demonstrates that high performance does not require massive scale. Key findings include:

  • Comparison with UNet: On the ISIC2018 dataset, MALUNet improved the Mean Intersection over Union (mIoU) by 2.39%. Remarkably, it achieved this while using 44x fewer parameters (0.175M vs. 7.77M) and 166x less computational complexity (0.083 GFLOPs vs. 13.78 GFLOPs).
  • Comparison with State-of-the-Art (SOTA): When compared against heavy models like TransFuse, MALUNet remains highly competitive. While TransFuse achieves a slightly higher mIoU (80.63% vs. 80.25%), MALUNet is 150x smaller in terms of parameters and 139x lighter in terms of GFLOPs.
  • Qualitative Analysis: Visual inspections of the segmentation masks show that MALUNet produces cleaner edges and fewer "hollow" artifacts than other lightweight models, such as UNeXt.

Conclusion and Future Work

MALUNet proves that through the intelligent design of attention mechanisms, it is possible to significantly reduce a model's footprint while actually improving its accuracy. By utilizing dataset-wide learning via IEA and multi-stage fusion via CAB and SAB, the architecture overcomes the traditional limitations of lightweight networks.

While currently optimized for skin lesion segmentation, the authors acknowledge that future research is needed to test MALUNet’s generalizability across other medical imaging modalities, such as CT or MRI scans. Future iterations may also incorporate Neural Architecture Search (NAS) and pruning to further refine the model's efficiency for the next generation of mobile medical diagnostics.

Summary updated

BibTeX
@inproceedings{ruan2022malunet,
  author = {Jiacheng Ruan and Suncheng Xiang and Mingye Xie and Ting Liu and Yuzhuo Fu},
  title = {{MALUNet: A Multi-Attention and Light-weight UNet for Skin Lesion Segmentation}},
  booktitle = {BIBM},
  year = {2022},
  doi = {10.1109/bibm55620.2022.9995040},
}
Figure