A Multi-Stage Faster RCNN-Based iSPLInception for Skin Disease Classification Using Novel Optimization.

Area of Science:

• Dermatology
• Computer Science
• Medical Imaging

Background:

• Skin cancer detection is crucial for patient outcomes.
• Deep learning (DL) methods, like Convolutional Neural Networks (CNNs), show promise but can overfit.
• Existing methods require improvement for accurate classification of benign and malignant skin tumors.

Purpose of the Study:

• To propose a novel multi-stage faster RCNN-based iSPLInception (MFRCNN-iSPLI) method for enhanced skin cancer classification.
• To address the overfitting problem inherent in traditional CNNs for skin lesion analysis.
• To improve the efficiency and accuracy of classifying both benign and malignant skin tumors.

Main Methods:

• The study introduces the MFRCNN-iSPLI method, integrating iSPLInception (based on Inception-ResNet) within a multi-stage classification framework.
• The prairie dog optimization algorithm is employed for candidate box deletion to refine classification.
• Performance was evaluated on the ISIC 2019 and HAM10000 skin lesion datasets.

Main Results:

• The MFRCNN-iSPLI method achieved high performance metrics: 95.82% accuracy, 96.85% precision, 96.52% recall, and 0.95% F1 score.
• The proposed method demonstrated superior effectiveness compared to existing techniques like CNN, hybrid DL, Inception v3, and VGG19.
• Analysis verified the method's prediction and classification capabilities in distinguishing skin lesions.

Conclusions:

• The MFRCNN-iSPLI method offers a robust and effective solution for skin cancer detection and classification.
• The multi-stage approach effectively mitigates overfitting and enhances diagnostic accuracy.
• This deep learning model shows significant potential for improving early skin cancer identification.