Automated Melanocytic Lesion Classification: Capsule Networks Trained With Synthetic Images Can Outperform Networks...

Area of Science:

• Dermatology
• Computer Science
• Medical Imaging

Background:

• Convolutional Neural Networks (CNNs) exhibit inherent classification errors.
• Capsule Networks (CNs), introduced in 2017, address many CNN shortcomings.
• Automated classification of melanocytic lesions is challenged by limited high-quality training data.

Purpose of the Study:

• To evaluate the classification performance of a Capsule Network (CN) on dermoscopic images of benign and atypical melanocytic lesions.
• To compare the efficacy of training a CN using real images versus synthetic images generated by its autoencoder.
• To assess the diagnostic capabilities of CNs in differentiating between benign and atypical melanocytic lesions.

Main Methods:

• A CN was developed and trained using 500 real dermoscopic images (250 benign, 250 atypical) from ISIC and PH2 datasets.
• The CN was subsequently trained on 3000 synthetic images generated via its autoencoder.
• Both CN models were tested on the same set of original real test images.

Main Results:

• The CN trained on real images achieved a diagnostic odds ratio (DOR) of 51.0.
• The CN trained on synthetic images achieved a significantly higher DOR of 71.3.
• The synthetic image-trained CN demonstrated superior performance compared to the real image-trained CN.

Conclusions:

• Training classifiers with synthetic images, which are easily generated in large quantities, offers inherent advantages for automated melanocytic lesion classification.
• CNs trained on extensive synthetic datasets show potential for unprecedented generalization performance in skin lesion classification.
• Capsule networks present a viable and advantageous alternative to CNNs for improving the accuracy of automated dermatological diagnostics.