Multi-level Attentive Skin Lesion Learning for Melanoma Classification.

Area of Science:

• Dermatology
• Medical Imaging
• Artificial Intelligence

Background:

• Melanoma classification is crucial for skin lesion diagnosis but is challenging due to variations in lesion appearance and imaging noise.
• Accurate melanoma detection is vital for patient outcomes and requires robust diagnostic tools.

Purpose of the Study:

• To propose a novel deep learning network, the multi-level attentive skin lesion learning (MASLL) network, to improve the accuracy of melanoma classification.
• To address the challenges posed by variable skin lesion appearances and dermoscopic imaging noise.

Main Methods:

• Developed a multi-level attentive skin lesion learning (MASLL) network incorporating a skin lesion localization (SLL) module for focused feature learning.
• Implemented a weighted feature integration (WFI) module to combine global and local features, enhancing discriminative capabilities.
• Evaluated the MASLL network on the ISIC 2017 dataset for melanoma classification.

Main Results:

• The proposed MASLL network demonstrated significant effectiveness in melanoma classification tasks.
• The integration of local learning via the SLL module and weighted feature fusion via the WFI module improved feature discriminability.
• Experimental results confirmed the superiority of the proposed method over existing approaches on the ISIC 2017 dataset.

Conclusions:

• The MASLL network offers an effective solution for enhancing melanoma classification in skin lesion diagnosis.
• The proposed SLL and WFI modules contribute to improved feature learning and fusion, leading to better classification performance.
• This research advances the application of AI in dermatology for more accurate and reliable skin cancer detection.