An auxiliary tool for preliminary tests of skin cancer : A self-modifying meta-learning method for clean and noisy data

Deep learning is a popular method when it comes to disease detection problems. As for skin cancer, a rather common kind of disease, however, unexpected rare cases often occur with few written records and referential materials, resulting in a disadvantageous situation for a usual neural network to learn. Hence, J in this paper, we propose a self-modifying meta-learning model which combines the idea of meta-learning with curriculum learning. Applying this mechanism, our model will first train on data of common diseases and then adapt the model to rare disease classification. Moreover, despite the existence of natural noise in data, like manually mistaken labels, our model can still handle which. We evaluate our algorithm on ISIC 2018 skin lesion classification dataset. Employing only 5 samples from each class, we achieve our accuracy up to 79.2%. Apart from that, when predicting data with a 20% noisy rate, our model can also adapt to classify unseen classes by accuracy of 76.2%. The further utilization for our model can be limited to skin cancer detection and diagnosis and extend to be applied to all kinds of diseases, serving as helpful assistants for medical workers, which would be a win-win for both patients and doctors.