SIL-Net: A Semi-Isotropic L-shaped network for dermoscopic image segmentation

Area of Science:

• Medical Imaging
• Computer Vision
• Artificial Intelligence

Background:

• Deep learning algorithms, particularly Convolutional Neural Networks (CNNs), are crucial for dermoscopic image segmentation in skin cancer detection.
• Traditional CNNs struggle with feature loss during down-sampling/up-sampling, impacting segmentation accuracy.
• Vision isotropic architectures offer improvements but are not directly applicable to segmentation tasks.

Purpose of the Study:

• To develop an efficient deep learning architecture that preserves the benefits of isotropic designs for clinical dermoscopic image segmentation.
• To address limitations in current CNNs for precise skin lesion boundary identification.

Main Methods:

• Introduction of the Semi-Isotropic L-shaped network (SIL-Net) for dermoscopic image segmentation.
• Development of a Patch Embedding Weak Correlation (PEWC) module to enhance inter-patch feature interaction.
• Integration of a Residual Spatial Mirror Information (RSMI) path to improve feature supplementation during up-sampling.
• Design of a Depth Separable Transpose Convolution (DSTC) based up-sampling module for refined feature reconstruction.

Main Results:

• SIL-Net achieved state-of-the-art performance on ISIC-2017, ISIC-2018, and PH² datasets, with Dice coefficients of 89.63%, 93.47%, and 95.11% respectively.
• Mean Intersection over Union (MIoU) scores reached 82.02%, 88.21%, and 90.81% on the same datasets.
• Demonstrated robustness and generalizability on intestinal polyp datasets (CVC-ClinicDB, Kvasir-SEG).

Conclusions:

• SIL-Net shows significant potential for precise skin lesion segmentation, meeting clinical diagnostic requirements.
• The proposed semi-isotropic design mechanism proves effective, leading to state-of-the-art performance across multiple datasets.
• The network exhibits strong generalizability and robustness, highlighting its clinical applicability.