Channel Attention Separable Convolution Network for Skin Lesion Segmentation

Changlu Guo,Jiangyan Dai,Marton Szemenyei,Yugen Yi

2023-09-03

Abstract:Skin cancer is a frequently occurring cancer in the human population, and it is very important to be able to diagnose malignant tumors in the body early. Lesion segmentation is crucial for monitoring the morphological changes of skin lesions, extracting features to localize and identify diseases to assist doctors in early diagnosis. Manual de-segmentation of dermoscopic images is error-prone and time-consuming, thus there is a pressing demand for precise and automated segmentation algorithms. Inspired by advanced mechanisms such as U-Net, DenseNet, Separable Convolution, Channel Attention, and Atrous Spatial Pyramid Pooling (ASPP), we propose a novel network called Channel Attention Separable Convolution Network (CASCN) for skin lesions segmentation. The proposed CASCN is evaluated on the PH2 dataset with limited images. Without excessive pre-/post-processing of images, CASCN achieves state-of-the-art performance on the PH2 dataset with Dice similarity coefficient of 0.9461 and accuracy of 0.9645.

Image and Video Processing,Computer Vision and Pattern Recognition

What problem does this paper attempt to address?

The paper aims to address the problem of skin lesion segmentation, particularly the need to automatically and accurately segment lesion areas from skin images for early diagnosis of skin cancer (such as melanoma). Manual segmentation of skin lesion images is time-consuming and prone to errors, thus there is an urgent need to develop precise and automated segmentation algorithms.

To tackle this challenge, the authors propose a new network architecture—Channel Attention Separable Convolutional Network (CASCN), which combines various advanced mechanisms and techniques, including U-Net structure, DenseNet, separable convolutions, channel attention mechanism, and Atrous Spatial Pyramid Pooling (ASPP). The main innovations of CASCN are:

1. **Utilizing DenseNet structure**: Reducing feature redundancy and improving feature propagation through dense blocks and transition blocks.

2. **Adopting separable convolutions**: Using depthwise separable convolutions in the decoder to create a more general and lightweight network, reducing computational complexity.

3. **Introducing an improved channel attention mechanism (MECA)**: Enhancing the network's discriminative ability and recovering spatial information lost due to pooling.

4. **Integrating the ASPP module**: Using ASPP between the encoder and decoder to capture multi-scale information, improving the model's recognition and segmentation accuracy for targets of different sizes.

Evaluated on the PH2 dataset, CASCN achieved state-of-the-art performance in terms of Dice similarity coefficient and accuracy, with scores of 0.9461 and 0.9645, respectively. Additionally, the paper explores the impact of key settings such as input image size, optimization methods, loss functions, and image augmentation on segmentation performance, and validates the effectiveness of each component through ablation experiments. Overall, CASCN demonstrates excellent performance in the task of skin lesion segmentation, showing promise in assisting doctors for more accurate early diagnosis.