ZooME: Efficient Melanoma Detection Using Zoom-in Attention and Metadata Embedding Deep Neural Network

Area of Science:

• Dermatology
• Medical Imaging
• Artificial Intelligence

Background:

• Melanoma detection is challenging for both clinicians and computer-aided diagnosis (CAD) systems.
• Increasingly complex deep learning models for melanoma classification have not consistently improved performance.
• There is a need to enhance the reliability of CAD for melanoma by incorporating clinically relevant data.

Purpose of the Study:

• To develop an improved computer-aided diagnosis network for melanoma detection.
• To enhance the extraction of pathological information from dermoscopy images.
• To integrate patient demographic data for more comprehensive melanoma prediction.

Main Methods:

• Proposed the Zoom-in Attention and Metadata Embedding (ZooME) network.
• Introduced a Zoom-in Attention model for detailed pathological feature extraction from dermoscopy images.
• Embedded patient metadata (age, gender, body site) into the network for enriched prediction.

Main Results:

• Achieved state-of-the-art performance on the ISIC-2020 dataset (33,126 images).
• Obtained an Area Under the Curve (AUC) score of 92.23%.
• Reported high diagnostic metrics: 84.59% accuracy, 85.95% sensitivity, and 84.63% specificity.

Conclusions:

• The ZooME network effectively improves melanoma detection by combining advanced image analysis with patient metadata.
• Integrating demographic information alongside dermoscopy images enhances the predictive power of CAD systems.
• The proposed approach offers a more credible and robust CAD solution for melanoma diagnosis.