SkinNet-INIO: Multiclass Skin Lesion Localization and Classification Using Fusion-Assisted Deep Neural Networks and...

Area of Science:

• Dermatology
• Computer Science
• Artificial Intelligence

Background:

• Deep convolutional neural networks (DCNNs) show promise in skin lesion classification but face challenges like limited data and class variability.
• Accurate classification is hindered by inter-class similarity, intra-class variation, and difficulty focusing on critical lesion areas.

Purpose of the Study:

• To develop an automated deep learning framework for multiclass skin lesion classification using dermoscopy images.
• To enhance contrast, optimize hyperparameters, fuse deep features, and select the best features for improved classification accuracy.

Main Methods:

• A novel preprocessing technique using dark channel haze and top-bottom filtering for contrast enhancement.
• Fine-tuning of three pre-trained DCNN models with transfer learning and hyperparameter optimization via a genetic algorithm (GA).
• Feature extraction, fusion using a serial correlation-based approach, and selection using an improved anti-Lion optimization algorithm.

Main Results:

• The framework achieved high accuracy, reaching 96.1% on the ISIC2018 dataset and 99.9% on the ISIC2019 dataset.
• The proposed method demonstrated improved accuracy compared to existing state-of-the-art techniques.
• Automated hyperparameter selection and feature selection/fusion processes enhanced learning and performance.

Conclusions:

• The framework effectively enhances contrast in cancerous regions, improving diagnostic capability.
• Automated techniques for hyperparameter and feature selection optimize the learning process and classification accuracy.
• The proposed fusion and selection methods maintain high accuracy while reducing computational time.