MiniDigPath: A New Standard For Pathology Images Few-Shot Learning Classification

Recently, few-shot learning(FSL) has received widespread attention. With the continuous development of machine learning and deep learning, the classification problem based on FSL has been deeply studied in fields such as computer vision and natural language processing. This greatly alleviates the problem of insufficient labeled data. So far, the publicly available datasets in FSL include Omniglot, miniImageNet, CUB-200, and CIFAR 100, etc. These datasets contain handwritten characters from different languages and natural images. No dataset in the medical field yet that can be applied to FSL despite massive of unlabeled data available. Therefore, the miniDigPath dataset is proposed by our work for FSL, which composed by 6 public pathology image datasets. It consists 67 different diseases and tissue types, and every type has 48-500 tissue image blocks. In total, there are 21165 histopathology images. Importantly, miniDigPath is available publicly for every researcher. It explores a new idea to solve pathological images annotation using FSL, which is the importance and originality of the dataset we proposed. Experimental evaluation on the classical FSL algorithm shows that the miniDigPath dataset can accomplish the task of FSL. Besides, FSL has good advantage for classification of digital pathology images.