Few-shot learning for segmentation of yeast cell microscopy images

Abstract

Cell segmentation from microscopic images can be performed using deep neural networks or image processing techniques. In addition to their inherent difficulties, these techniques come together with the requirement of feeding the neural network with a large number of image samples in order to obtain a good result. However, this is not sustainable in terms of collecting and labeling microscopic images and represents a costly and time-consuming solution for every new microscopic image and cell type. Instead, fine-tuning can be employed by taking advantage of the adaptation ability of a model trained using meta-learning algorithms. In this way, while more general and better results can be obtained with fewer samples, the training process does not start from scratch for each new cell type or data set. In this article, microscopic images of yeast cells were recorded and analyzed using Reptile algorithm. After fine-tuning with a small number of samples, an average success rate of 81 % IoU (Intersection over Union) was obtained on the test pictures in addition to the model accuracy reaching up to 87%.