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

BACKGROUND/OBJECTIVES: Convolutional neural networks (CNNs) are known, due to inherent flaws in their design, to be subject to classification error. Many of these shortcomings in classification performance were addressed in 2017 with the introduction of capsule networks (CNs). The objective of this investigation is to determine the classification performance of a CN with respect to dermoscopic images of benign and atypical melanocytic lesions. We implement two kinds of training data for comparison: first, real images; and second, by utilising the CNs' autoencoder, a large number of synthetic images. METHODS: 500 melanocytic lesion images (250 benign and 250 sufficiently atypical to warrant excision) were obtained from ISIC and PH2 Datasets. Images were randomly split equally for training and testing, and equally split between the classification categories. A CN was developed, trained and tested on these data. The CN was then trained on 3000 synthetic images (with an equal number of images in each classification category). The newly trained CN was then tested on the original test images. RESULTS: The CN trained on the real images yielded a DOR (diagnostic odds ratio) of 51.0, a result which is comparable and, in many cases, superior to those reported for CNNs. The same CN trained on the synthetic images yielded a DOR of 71.3, a result which is significantly higher than the result obtained with real-image training. CONCLUSIONS: Since the effectiveness of automated melanocytic lesion image classification is limited by the difficulty in obtaining large numbers of high-quality training data, our results support the idea that training a classifier utilising synthetic images-where large numbers of quality images can be easily generated-is inherently advantageous. Given its advantages over a CNN, and trained on a very large number of synthetic images of melanocytic lesions, a CN classifier thus has the potential, only limited by computer resources, to yield previously unseen levels of generalisation performance.