Modified EfficientNet-B0 Architecture Optimized with Quantum-Behaved Algorithm for Skin Cancer Lesion Assessment.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Computational Biology

Background:

• Skin cancer is a prevalent global disease with significantly improved survival rates when detected early.
• Late diagnosis of skin cancer drastically increases mortality risk, highlighting the need for accurate and efficient detection methods.

Purpose of the Study:

• To develop and validate a novel deep learning framework for enhanced early and accurate detection of skin cancer.
• To improve diagnostic accuracy and assist dermatologists in clinical decision-making for better patient outcomes.

Main Methods:

• A modified EfficientNet-B0 architecture incorporating Mobile Inverted Bottleneck Convolution with Squeeze and Excitation was developed.
• The model utilizes both 3x3 and 5x5 kernels for balanced feature extraction and increased learning capacity.
• Hyperparameters and feature vectors were optimized using the Quantum-Behaved Particle Swarm Optimization (QBPSO) algorithm on benchmark dermoscopic image datasets (HAM10000, ISIC 2019, MSLD v2.0).

Main Results:

• The proposed framework achieved an average accuracy (mAcc) of 99.62% on the HAM10000 dataset and 92.5% on the ISIC2019 dataset.
• The model demonstrated superior performance compared to various state-of-the-art models and techniques.
• Reliability and stability were confirmed through Monte Carlo simulations, indicating robust performance.

Conclusions:

• The developed framework effectively reduces skin cancer diagnostic errors.
• It serves as a valuable tool to support dermatologists in clinical decisions, ultimately leading to improved patient outcomes.
• The approach addresses challenges such as data imbalance and interpretability in medical image analysis.