Melanoma Recognition by Fusing Convolutional Blocks and Dynamic Routing between Capsules.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Dermatology

Background:

• Skin cancer, particularly melanoma, is a significant global health concern.
• Accurate and early diagnosis is crucial for effective melanoma treatment.
• Current deep learning models like CNNs face limitations in handling spatial hierarchies and transformations.

Purpose of the Study:

• To propose a novel deep learning architecture for enhanced automatic melanoma diagnosis.
• To overcome limitations of traditional Convolutional Neural Networks (CNNs) in analyzing skin lesion images.
• To improve the extraction of abstract features from high-resolution skin lesion images.

Main Methods:

• Developed a hybrid architecture integrating convolutional blocks with a customized Capsule Network (CapsNet).
• Utilized high-quality 299x299x3 skin lesion images for training and validation.
• Performed hyper-parameter tuning to optimize learning with limited training data.
• Validated model predictions using two modern model-agnostic interpretation tools.

Main Results:

• The proposed architecture significantly outperformed several state-of-the-art models across eleven diverse image datasets.
• Demonstrated superior performance in extracting richer abstract features compared to conventional methods.
• The model showed effective learning capabilities even with constrained training datasets.

Conclusions:

• The hybrid CNN-CapsNet architecture offers a promising advancement for automatic melanoma diagnosis.
• This approach effectively addresses the limitations of existing deep learning models in medical image analysis.
• The findings suggest potential for improved clinical decision support in dermatology.