Conditional Generative Adversarial Network for Mitigating Skin Diseases Identification Modeling with Limited Data

Accurate skin diseases prediction modelling has always been a challenging task. On of the contributing factor is due to the limited number of its corresponding medical images. Various traditional images augmentation such as cropping, flipping, rotation and bright manipulation are normally implemented on pre-processing to improve the dataset. However, those approaches are less significant to improve the model’s performance since duplication information and less variation led to bad generalization. This study presents the adoption of Conditional Generative Adversarial Networks (CGAN) for constructing additional synthetical skin diseases image dataset to improve the prediction model’s performance. HAM10000 and Fitzpatrick 17k skin diseases dataset are used as the use case. The pre-trained model, VGG16, is applied as the prediction modelling. Based on the simulation results on given datasets, the attached CGAN procedure contributes to the improvement especially on precision. It indicates that the additional synthetical dataset produced by CGAN support model on more sensitive to false positive case.