Using Synthesized Data to Train Deep Neural Net with Few Data

Abstract

As Computer-Assisted Surgery (CAS) getting popular, more and more research has been conducted to help surgeons operate. We aim at the semantic segmentation in the endoscopy surgery scenario because semantic segmentation is the first step for a computer to grasp what shows up in the vision of an endoscope. However, modern Deep Learning algorithms need myriads of training data. Since data of the endoscopy surgery scene is relatively scarce, the performance of existing algorithms is thus rather limited. Therefore, we tried to solve the problem of training a semantic segmentation network with few data in this work. We propose a proof-of-concept system offering the ability to enlarge the dataset and improve the performance. The system aims to synthesize a pair of training data in a single pass and provides a sufficient amount of data to train a network. We evaluated our method using the dataset provided by MICCAI 2018 Robotic Scene Segmentation Sub-Challenge. Our method yielded 11.79% mIoU improvement in recognizing anatomical objects and 2.2% mIoU in recognizing surgical instruments. Recognizing anatomical objects accurately would definitely benefit CAS. Preliminary results suggest our method helps the classifier become more robust and accurate even if not having large amount of data.