Rema-Net: An efficient multi-attention convolutional neural network for rapid skin lesion segmentation

Area of Science:

• Dermatology
• Medical Imaging
• Computer Vision

Background:

• Accurate skin lesion segmentation from dermoscopic images is crucial for clinical diagnosis and treatment.
• Convolutional neural networks (CNNs), like U-Net, are prevalent for skin lesion segmentation but often require substantial computational resources and time.
• Existing CNN models can be resource-intensive, hindering rapid training and segmentation applications.

Purpose of the Study:

• To develop an efficient and effective convolutional neural network (Rema-Net) for rapid skin lesion segmentation.
• To reduce the computational complexity and training time of skin lesion segmentation models.
• To enhance the accuracy of skin lesion segmentation using a novel multi-attention mechanism.

Main Methods:

• Proposed Rema-Net, an efficient multi-attention convolutional neural network designed for fast skin lesion segmentation.
• Implemented a simplified down-sampling module using convolutional and pooling layers with spatial attention.
• Incorporated reverse attention operations on skip-connections to improve segmentation performance.

Main Results:

• Rema-Net demonstrated a reduction of nearly 40% in parameters compared to U-Net.
• Achieved significantly improved segmentation metrics over previous methods on five public datasets (ISIC-2016, ISIC-2017, ISIC-2018, PH2, HAM10000).
• Generated segmentation predictions that more closely match the actual lesion boundaries.

Conclusions:

• Rema-Net provides an efficient and accurate solution for rapid skin lesion segmentation.
• The proposed network architecture effectively balances performance and computational efficiency.
• This method has the potential to facilitate faster and more accessible skin lesion analysis in clinical settings.