Deep Transfer Learning Based Web Interfaces for Biology Image Data Classification

Abstract

Artificial Intelligence (AI) continues to play an integral role in the modernization of the medical and scientific fields. Recent advances in machine learning such as the Convolutional Neural Network (CNN), have demonstrated the ability to recognize complex images with very low error rates using relatively small data set (thousands of images) to become fully trained. Scientists including who are not familiar with programming begin to recognize the need to incorporate machine learning in their research methods to improve the accuracy and the speed of diverse data manipulation without depending on computer scientists. Several tools are developed to serve for these purposes, but such tools are mostly targeting data scientists and often too general or too many options to configure for biologists without machine learning knowledge to get started. We present our work on incorporating Deep Learning into one specific research pipeline that studies how genes work together to regulate muscle formation in the vertebrate frog embryo, Xenopus laevis. This research method uses a knockdown approach to diminish the expression of key genes and study how this loss-of-gene function affects the process of muscle formation and differentiation, using mostly fluorescent microscopy techniques which requires time-consuming and challenging visual classifications. We utilized CNN-pretrained transfer learning on the data set with a few different hyper parameters and trained a model with 99% accuracy. Using this experience and discussion with scientists new to machine learning, we developed web interfaces for easy-to-use and complete workflow for scientists to create different classification classes to train, predict and incorporate into their research pipeline.