A study on skin tumor classification based on dense convolutional networks with fused metadata

Area of Science:

• Dermatology
• Artificial Intelligence
• Medical Imaging

Background:

• Skin cancer diagnosis accuracy is suboptimal, with rates below 80% for experienced dermatologists and 60% for inexperienced ones.
• Misdiagnosis can lead to delayed treatment, negatively impacting patient prognosis.
• Existing neural network models primarily use image data, neglecting valuable clinical metadata.

Purpose of the Study:

• To develop a deep convolutional neural network model for enhanced skin cancer classification.
• To improve diagnostic accuracy by integrating patient clinical metadata with image features.
• To address limitations in current AI-based skin cancer detection methods.

Main Methods:

• Utilized a pre-trained DenseNet-169 model to extract high-level image features (edges, color, texture, form).
• Introduced MetaNet and MetaBlock modules to incorporate and weight clinical metadata, refining feature extraction.
• Combined features from MetaNet and MetaBlock into the MD-Net module for classification.

Main Results:

• The proposed MD-Net model achieved a balancing accuracy of 81.4% on the PAD-UFES-20 and ISIC 2019 datasets.
• Demonstrated an 8% to 15.6% increase in diagnostic accuracy compared to previous methods.
• Successfully improved classification for challenging cases like actinic keratosis and skin fibromas.

Conclusions:

• Integrating clinical metadata with deep convolutional networks significantly enhances skin cancer diagnostic accuracy.
• The developed MD-Net model offers a promising advancement in AI-driven dermatological diagnostics.
• This approach has the potential to reduce misdiagnosis and improve patient treatment timelines.