Fusion of Local and Global Feature Representation With Sparse Autoencoder for Improved Melanoma Classification.

Area of Science:

• Dermatology
• Medical Imaging
• Artificial Intelligence

Background:

• Automated skin cancer diagnosis faces challenges like image variations and detail loss in Convolutional Neural Networks (CNNs).
• Existing methods often fail to capture both local and global features from dermoscopy images effectively.
• Downsampling high-resolution images in CNNs can lead to the loss of crucial diagnostic details.

Purpose of the Study:

• To propose an advanced framework for melanoma classification that overcomes limitations of existing automated diagnostic methods.
• To enhance the accuracy of skin cancer diagnosis by preserving and utilizing both local and global image features.
• To improve the representation of image features for more robust melanoma classification.

Main Methods:

• A novel framework employing ensemble feature fusion and a sparse autoencoder (SAE) is proposed.
• Features are extracted from two streams: local (image patches) and global (whole image) using a pre-trained CNN.
• Fused features undergo further enrichment using the sparse autoencoder (SAE) framework.

Main Results:

• The proposed method demonstrated superior performance in melanoma classification on the ISIC 2016 dataset.
• Combining local and global features effectively addressed the issue of detail loss from image downsampling.
• The sparse autoencoder (SAE) successfully enriched feature representation, leading to improved diagnostic accuracy.

Conclusions:

• The ensemble feature fusion and SAE-based framework significantly enhances automated melanoma classification.
• This approach offers a more comprehensive analysis of dermoscopy images, improving diagnostic accuracy.
• The method provides a promising advancement for computer-aided diagnosis in dermatology.