Automatic skin lesion classification using a new densely connected convolutional network with an SF module.

Area of Science:

• Artificial Intelligence
• Medical Imaging
• Computer Vision

Background:

• Automatic skin lesion classification from dermoscopy images is difficult due to diverse lesion appearances, image artifacts, limited data, and current method limitations.
• Existing approaches struggle with the complex visual characteristics of skin lesions.

Purpose of the Study:

• To develop an advanced deep learning framework for accurate skin lesion classification.
• To enhance the feature extraction capabilities for improved diagnostic accuracy in dermatology.

Main Methods:

• A novel densely connected convolutional network, DenseSFNet-45, was developed by integrating an SE-Fire (SF) block into a DenseNet architecture.
• A two-stage framework was proposed, involving initial skin lesion segmentation followed by classification on the segmented region.
• The SF block combines Fire modules and squeeze-and-excitation (SE) blocks to leverage spatial and channel-wise information.

Main Results:

• The proposed DenseSFNet-45 model demonstrated superior performance compared to traditional machine learning algorithms, classical models, baselines, and state-of-the-art methods.
• The two-stage segmentation-then-classification approach enabled the network to extract more specific and discriminative features.
• Extensive evaluations on ISBI-skin-2017, ISBI-skin-2018, and PH2 datasets confirmed the method's effectiveness.

Conclusions:

• The developed DenseSFNet-45 and the two-stage framework offer a significant advancement in automated skin lesion classification.
• This approach holds promise for improving the accuracy and efficiency of diagnosing skin conditions from dermoscopy images.
• The method's superior performance highlights the potential of integrating advanced deep learning architectures for medical image analysis.