Image Recognition Technology Using Deep Learning

For radiation therapy of tumors that move with respiration, tumor-tracking radiotherapy is available through high-precision irradiation that can be achieved by inserting a metal marker adjacent to a tumor and identifying the marker’s position using X-ray fluoroscopic images. There are also ways to use the marker’s position in CT images for patient positioning and therapy. In any case, the marker’s position in X-ray images must be identified using image recognition technology, but that may be difficult with the conventional template matching method if a non-spherical marker is used, because the image changes during motion.

Our laboratory focuses on a high-precision marker detection technique using deep learning by building a neural network to classify a large number of marker images and background images and using the output classification scores to determine the marker position. Since it is difficult to obtain marker images of all shapes with various background images during experiments, we are also studying ways to produce the necessary imaging data for learning based on digital data on the human body.