Intracellular Strain Measurement of Oocyte Based on Inclusion Appearance Using Neural Network

Somatic cell nuclear transfer(SCNT) involves many important procedures, of which enucleation is a critical one. During the extraction step of enucleation, oocytes undergo large deformation. Because of the similarity of appearance throughout the oocytes and the 3D spatial property of oocytes, traditional deformation measurement methods can hardly be used in full-field strain measurement during this process, especially for realtime control. Neural networks have the properties of large-scale parallel processing, high redundancy, and nonlinear operations. They can achieve high computing speed while maintaining high accuracy and strong fault tolerance. But lack of labeled data in different situations is the main problem. In this paper, a2D dataset for training Recurrent All-Pairs Field Transforms (RAFT) network was constructed. The dataset simulates the characteristics of oocytes. By fine-tuning a pre-trained model on the dataset, the resulting network outperforms the pre-trained model and traditional optical flow methods. At the same time, with GPU acceleration, it can deal with 640*480 resolution images at a speed of 4fps, which is comparable to traditional methods.