FEW-SHOT TRANSFER LEARNING FOR BRAIN TUMOR CLASSIFICATION

Abstract

Analysis of MRI images to detect the type of tumor in the human brain is a field of intensive research where AI researchers strive to provide fast and robust tumor detection solutions, to support medical professionals. Considering the limited resources at hand, the aim is to maximize efficiency so as to create a deployable system that can be trained using a limited number of scanned images.   Transfer learning can be used to classify MRI images by reusing the knowledge from another similar domain. However, it does not perform well when new task is different than the old one. Few-shot learning reduces the dependency on task similarity. In this work, we have used Siamese networks to perform Few-Shot Learning with Transfer Learning for Brain Tumor Classification, empowering the model to generalize well with a small number of labeled samples. The model after implementation gave us an accuracy of 85.05% with a precision value of 92.06% and a recall value of 87.87%.