ESTRO 2023 - Abstract Book

S1397

Digital Posters

ESTRO 2023

Conclusion Great Vessels area is a proper guide for organ based deformation, while MDA combined with visual inspection are most suitable methodology to evaluate DIR in pancreatic recurrence back propagation to initial CT/MR.

PO-1687 Iterative reconstruction of MVCT with deep neural networks

S. Ozaki 1 , S. Kaji 2 , K. Nawa 3 , T. Imae 3 , K. Nakagawa 3

1 Hirosaki University, Graduate School of Science and Technology, Hirosaki, Japan; 2 Institute of Mathematics for Industry, Kyushu University, Division for Intelligent Societal Implementation of Mathmatical Computation, Fukuoka, Japan; 3 University of Tokyo Hospital, Department of Radiology, Tokyo, Japan Purpose or Objective Deep learning is widely used for image quality enhancement in medical imaging. However, deep learning-based methods usually need a large amount of training data, and collecting such big data is often expensive for the medical staff. Deep image prior (DIP) gives one of the solutions; it can improve image quality without training data. By extending the DIP, we propose a novel method of CT reconstruction with deep neural networks. We apply our methods to MVCT reconstruction, which is used in helical tomotherapy. Materials and Methods Iterative reconstruction (IR) solves the inverse problem by minimizing a loss function. Often regularization terms such as the total variation (TV) are introduced to incorporate prior on the images. Instead of optimizing the voxel values directly, the DIP can optimize the parameters of a convolutional neural network (CNN) whose output is the voxel values. This indirect optimization can be thought of as an application of DIP for CT reconstruction, where the output image and the observed projection are taken in a consistency loss function. Our primary contribution is an optimization methodology for the DIP reconstruction where we start with large weights of low-resolution layer terms in the loss function, and then progressively increase the weights of high-resolution layer terms with an increasing iteration, as visualized in Figure 1. Our method stabilizes the optimization of model parameters, and improves high-resolution image quality.