Few-Shot Learning on Histopathology Image Classification
Ankit Kumar Titoriya, Maheshwari Prasad Singh · 2022
Cancer detection in histopathology slides is not easy even today. CNN (Convolutional Neural Network) based Object identification and segmentation algorithms work very well. A large dataset of medical images is required for classification which may not be available especially for rare diseases. Therefore, deep learning and machine learning may not be effective for rare disease classification. If CNN architecture is trained on one dataset then it performs well but the same architecture may not achieve good accuracy on other datasets. So, generalization is one of main issues. This paper proposes FSL (Few-Shot Learning) to solve generalization and size of dataset. This paper uses Prototypical networks and MAML (Model Agnostic Meta Learning) simultaneously on four different datasets. Along with this, it has also been checked whether these two networks meet the concept of generalization or not. The paper also finds accuracy of both networks in 2-way, 3-way, and 5-way modes. Simulation results show that MAML achieves accuracy of 84.56% in the 2-way 2-shot 2 query mode. Further, simulation results show that Prototypical Network achieves accuracy of 74.575%, 61.9889% and 45.762% in 2-way 2-shot 2 query mode, 3-way 3-shot 3-query mode and 5-way 5-shot 5-query mode, respectively.