基于深度特征学习的快速低剂量CT成像

本研究面向解决当前国内医院影像诊疗科室CT扫描中的亟待解决的辐射剂量问题，通过和临床影像科医生和企业工程师紧密合作，本项目将从深度特征学习的角度入手设计低剂量CT成像算法，从1)-深度卷积反投影网络; (2)-投影到图像的全网络; (3)-跳跃式迭代重建映射网络；(4)-组织区分性网络四种的深度卷积网络四个策略入手，充分挖掘不同剂量下的不同质量的CT 投影数据及图像数据中的区别性特征信息，开发的算法模型将具有对图像低、中、高级特征强大的学习表达能力。随后，课题组将在临床影像科医生和企业工程师的指导下完善和加速算法，实现临床实用的快速低剂量CT成像算法, 并将成像算法集成植入国内大型医疗设备生产企业的CT整机系统。开发的算法模型将不依赖于特定品牌的CT设备，具有较高临床可移植性重建算法。通过该项目的实施，课题组有望获得有至少2 项技术达到同类技术的国际领先水平并获得国际专利。

This study aims toward to provide feasible solution to urgent CT radiation problem in radioliogical departments in local hospitals. This project is to be implemented in close collaboration with clinical radiologists to develop low dose CT imaging algorithms from the following four deep convolution network strategies: 1)-Deep Convolution Back-projection Network; 2)-Complete Network from Projections to Images 3)-Jump-mode Iteration Reconstruction Mapping Network; 4)- Tissue Discriminative Networks。In this way, the discriminative feature information in CT data (including projection data and image data) in the different doses and qualities will be fully exploited, and the developped algorithms will have good ability in representing image features in all the low, medium and high levels. After this, the research group will continue accelerating and improving the algorithms with clinical radiologists and industrial engineers. The finalized algorithm will be integrated into the CT systems for the domestic companies producing large medical equipments. The application of the developed algorithms will not be limited to the specific brands and models for CT systems, and have good implantability. The developped imaging algorithms will have the potential to run up to internationally high level and result in at least two international patents.