Multi-level Context Aggregation Network for Skin Lesion Segmentation

The automated segmentation of skin lesions constitutes a crucial phase in the computer-assisted diagnosis of skin cancer, providing a foundation for further analysis. With the advancement in convolutional neural networks, the segmentation effectiveness of skin lesions has experienced marked enhancement. However, challenges still exist in skin lesion segmentation, such as variations in lesion size, irregular lesion shapes, and fuzzy boundaries between lesions and the background. To achieve efficient and precise segmentation of skin lesions, we present the Multi-level Context Aggregation Network (MLCANet). In the encoder section, we adopt an InceptionNeXt network based on large kernel convolutions to expand the receptive field and enhance the extraction capability of contextual information. In the skip connection section, we introduce the Gated Fully Fusion method to effectively reconstruct skip connections, aggregating multi-level contextual information. This method leverages the advantages of different layer feature maps to strengthen the encoder features. Additionally, we propose a residual multi-scale channel attention module at the end of the model to emphasize critical channel information, thereby enhancing the performance of the model. Our experiments on the publicly ISIC 2018 skin lesion dataset reveal that our approach surpasses other methods across five commonly employed evaluation metrics, affirming the outstanding segmentation performance of our approach.