Enhancing skin lesion classification: a CNN approach with human baseline comparison.

Area of Science:

• Dermatology
• Medical Imaging
• Artificial Intelligence

Background:

• Accurate diagnosis of malignant skin lesions is crucial for effective treatment.
• Variability in medical image analysis can impact diagnostic accuracy.
• Integrating artificial intelligence with human expertise offers potential improvements.

Purpose of the Study:

• To develop and evaluate an augmented hybrid approach for diagnosing malignant skin lesions.
• To enhance diagnostic accuracy by combining Convolutional Neural Network (CNN) predictions with selective human interventions.
• To assess the performance and resource efficiency of the hybrid approach compared to standalone methods.

Main Methods:

• An EfficientNetB3-based CNN was trained on ISIC-2019 and ISIC-2020 datasets.
• A hybrid approach was implemented, using high-confidence CNN predictions and expert human assessments for low-confidence predictions.
• Performance was evaluated on a 150-image test set using ROC curves, AUC, and analysis of human resource costs.

Main Results:

• The baseline CNN achieved an Area Under Curve (AUC) of 0.822.
• The augmented hybrid approach improved the true positive rate to 0.782 and reduced the false positive rate to 0.182.
• The hybrid approach demonstrated better diagnostic performance with minimal human involvement and analyzed human resource costs.

Conclusions:

• The augmented hybrid approach effectively combines CNNs and human expertise for improved skin lesion diagnosis.
• This method offers a scalable and resource-efficient solution for medical image analysis.
• The findings highlight the complementary strengths of AI and expert clinicians in dermatological diagnostics.