Fully transformer network for skin lesion analysis.

Area of Science:

• Dermatology
• Medical Imaging
• Artificial Intelligence

Background:

• Automatic skin lesion analysis, including segmentation and classification, is crucial but challenging due to data variability from diverse sources.
• Convolutional Neural Networks (CNNs) are widely used but struggle with capturing long-range dependencies in skin lesion images.
• The intrinsic locality of CNNs limits their ability to process global contextual information.

Purpose of the Study:

• To introduce a novel Fully Transformer Network (FTN) designed to enhance skin lesion analysis by effectively learning long-range contextual information.
• To address the limitations of CNNs in capturing global dependencies for improved segmentation and classification of skin lesions.
• To develop a computationally efficient model for skin lesion analysis.

Main Methods:

• Proposed a Fully Transformer Network (FTN) incorporating a hierarchical structure.
• Introduced the Spatial Pyramid Transformer (SPT) module within the FTN, integrating spatial pyramid pooling (SPP) with multi-head attention (MHA).
• Evaluated the FTN's performance on the ISIC 2018 dataset for skin lesion analysis tasks.

Main Results:

• The FTN demonstrated superior performance compared to state-of-the-art CNNs in skin lesion analysis tasks.
• FTN achieved higher computational efficiency and required fewer tunable parameters than existing CNN models.
• The proposed Spatial Pyramid Transformer (SPT) module significantly reduced computational and memory usage.

Conclusions:

• The Fully Transformer Network (FTN) offers an effective and efficient solution for automatic skin lesion analysis.
• FTN's ability to capture long-range contextual information surpasses the capabilities of traditional CNNs.
• The developed model provides a promising advancement for dermatological diagnostics through improved image analysis.