Enhancing Melanoma Diagnosis in Histopathology with Deep Learning and Synthetic Data Augmentation

Area of Science:

• Computational pathology
• Artificial intelligence in medicine
• Biomedical image analysis

Background:

• Accurate melanoma diagnosis from H&E images is hindered by limited and imbalanced datasets.
• Deep learning models struggle with data scarcity, impacting diagnostic performance.

Purpose of the Study:

• To investigate the effect of synthetic image generation using GANs on training deep learning classifiers for melanoma diagnosis.
• To evaluate the utility of synthetic data for balancing imbalanced biomedical datasets.

Main Methods:

• Trained ResNet-18 classifiers using real images and a mix of real and synthetic images (melanocytic nevus class).
• Evaluated models at resolutions up to 1024 × 1024 pixels.
• Assessed image quality using Fréchet Inception Distance (FID) and classification performance using standard metrics.

Main Results:

• Mixed models achieved competitive performance, particularly in specificity and reducing false negatives, compared to real-only models.
• The best-performing model (1024 × 1024 px, 50 epochs, mixed dataset) showed 96.00% accuracy and 97.00% specificity.
• False negatives were reduced from 80 to 75 cases with synthetic data augmentation.

Conclusions:

• Synthetic data generation via GANs can serve as a valuable complementary tool when real sample acquisition is limited.
• This approach shows potential for improving clinically relevant outcomes, especially in reducing missed melanoma diagnoses.
• Further research into conditional generation and synthesis of malignant samples is warranted.