Optimized Few-Shot Learning with Adversarial Feature Hallucination Networks for Multiclass Skin Disease Prediction Modeling

Deep Learning (DL) is a popular learning method known for its effectiveness in classifying data and its ability to handle large input datasets. However, in real-world applications, the amount of data recorded and understood by computers may not always be abundant or sufficient especially on medical data. To address this issue, the Few-Shot Learning (FSL) technique can serve as a solution, as it enables classification by learning from only a few labeled samples. Dealing with the case of prediction modeling for skin’s diseases identification, FSL has potential on addressing its limitation of valid annotated data. This paper presents the study of FSL potential on skin’s diseases prediction modeling by adopting Adversarial Feature Hallucination Networks (AFHN) to optimize the FSL proposed framework performance. This study conducts a comparative analysis of several setups on the ISIC 2019 dataset to demonstrate the performance differences and advantages of the proposed model. The final results indicate an improvement in performance subject to AFHN optimization. The results showed that the model with feature synthesis achieved an accuracy of 76.99%, and this result was further supported by the evaluation metrics.