ELK-BiONet: Efficient Large-Kernel Convolution Enhanced Recurrent Bidirectional Connection Encoding and Decoding Structure for Skin Lesions Segmentation

Abstract

The size and shape of skin lesions often exhibit significant variability, and enabling deep learning networks to adapt to this variability is crucial for improving the segmentation performance of such lesions. The encoder-decoder architecture has become one of the most commonly used structures for semantic segmentation in deep learning models. However, when the convolution-based UNet network is applied to skin lesion segmentation, several issues remain. (1) Traditional small-kernel convolutions have a limited receptive field, which makes it difficult to adapt to the varying sizes and shapes of skin lesions. (2) The conventional U-Net architecture experiences a substantial increase in parameter count as the network depth grows. (3) Although the U-Net decoder utilizes feature information from the encoder, the features extracted by the decoder are not fully leveraged. To address the above challenges in U-Net for skin lesion segmentation tasks, we propose an efficient large-kernel convolution enhanced recurrent bidirectional connection encoding and decoding structure for skin lesions segmentation (ELK-BiONet). The main innovations of this method are as follows: (1) We propose a large-kernel convolution method that balances large and small receptive fields while maintaining a relatively low parameter count. (2) The network extracts feature information in a recurrent manner, allowing the construction of deeper network architectures while keeping the overall parameter count nearly constant. (3) By employing bidirectional connections, the features extracted by the decoder are fully utilized in the encoder, thereby enhancing the segmentation performance of the network. We evaluated our method on skin lesion segmentation tasks, and the results demonstrate that our ELK-BiONet significantly outperforms other segmentation methods.