Multi-skin disease classification using hybrid deep learning model

Area of Science:

• Oncology
• Medical Imaging
• Computer Science

Background:

• Early detection and classification of skin cancer are crucial for improving patient survival rates.
• Computer-Aided Diagnosis (CAD) systems, particularly those utilizing Deep Learning (DL), are vital for assisting radiologists.
• Existing methods often require effective skin lesion segmentation for improved classification.

Purpose of the Study:

• To develop an advanced CAD system for early and accurate skin cancer detection and classification.
• To enhance skin lesion segmentation and classification performance using novel optimization techniques.
• To provide a competitive and effective tool for dermatologists in diagnosing skin cancer.

Main Methods:

• Preprocessing involved noise reduction using Adaptive Wiener Filter (AWF) and hair removal via Maximum Gradient Intensity (MGI).
• Skin lesion segmentation was performed using an optimized Region Growing (RG) method integrated with the Modified Honey Badger Optimiser (MHBO).
• Classification of skin cancer types was achieved using the Deep Learning model MobileSkinNetV2 on the ISIC dataset.

Main Results:

• The proposed system achieved high accuracy (99.01%) and precision (98.6%) in skin cancer classification.
• The integration of MHBO with RG significantly improved the segmentation process.
• Experimental results demonstrated competitive performance compared to existing skin cancer detection models.

Conclusions:

• The developed CAD system shows significant potential for early and accurate skin cancer diagnosis.
• The novel segmentation and classification approach offers a promising advancement in dermatological imaging analysis.
• This research provides valuable support for dermatologists in the fight against skin cancer.