Double-Condensing Attention Condenser: Leveraging Attention in Deep Learning to Detect Skin Cancer from Skin Lesion...

Area of Science:

• Artificial Intelligence
• Medical Imaging
• Oncology

Background:

• Skin cancer is the most prevalent cancer in the US, affecting 1 in 5 Americans.
• Current advanced skin cancer detection methods use complex deep neural networks, demanding significant computational resources.
• TinyML applications are exploring efficient architectures like Double-Condensing Attention Condensers (DC-AC) for faster computation.

Purpose of the Study:

• To develop an efficient deep neural network for skin cancer detection using DC-AC.
• To customize a self-attention neural network with DC-AC for analyzing skin lesion images.
• To achieve high performance in skin cancer detection with reduced computational complexity.

Main Methods:

• Designed a novel deep neural network incorporating DC-AC into a self-attention backbone.
• Trained and evaluated the model on the ISIC 2020 dataset for skin lesion classification.
• Assessed model performance using Area Under the Receiver Operating Characteristic Curve (AUC).

Main Results:

• The proposed DC-AC model achieved an AUC of 0.90 on the public ISIC 2020 test set and 0.89 on the private test set.
• The model utilizes only 1.6 million parameters and 0.32 GFLOPs, demonstrating high efficiency.
• Outperformed the Cancer-Net SCa network by over 0.13 in AUC.

Conclusions:

• The developed DC-AC network offers a computationally efficient and accurate solution for skin cancer detection.
• This approach represents a significant advancement over traditional, resource-intensive methods.
• The model is publicly available to support clinical decision-making and further research in AI for cancer detection.