摘要
肺结节检测的深度学习方法一般分为候选结节检测和假阳性结节消除两个阶段.基于两阶段方法,提出了一种整合新数据以提升系统准确性的增量学习加速方案.利用历史数据的训练模型对新数据进行筛选,把表现性能不好的数据作为继续训练两阶段模型的输入.在LUNA16和TIANCHI17两个经典数据集上对上述方法进行测试,只需利用一半以下的新训练数据就能取得与传统两阶段方法相同的效果.
The deep learning method for pulmonary nodule detection is generally divided into two stages: candidate nodule detection and false positive nodule elimination. Based on the two-stage method, an incremental learning acceleration scheme is proposed that integrates new data to improve the accuracy of the system. The training model of historical data screens new data and selects the data with poor performance as an input for the continuous training of the two-stage model. The above methods are tested on LUNA16 and TIANCHI17 two classic data sets. Using only half of the new ones, the new model can achieve the same e?ect as the traditional two-stage method.
作者
李正
胡贤良
梁克维
虞钉钉
LI Zheng, HU Xian-liang, LIANG Ke-wei, YU Ding-ding(School of Mathematical Sciences, Zhejiang University, Hangzhou 310027, Chin)
出处
《高校应用数学学报(A辑)》
CSCD
北大核心
2018年第2期127-139,共13页
Applied Mathematics A Journal of Chinese Universities(Ser.A)
基金
国家自然科学基金面上项目(11471253)
关键词
肺结节检测
深度学习
两阶段方法
假阳性
增量学习
pulmonary nodule detection
deep learning
two-stage method
false positives
incre-mental learning