loading page

Automatic CNS diseases real-time detection in first-trimester fetal ultrasound image via deep neural networks
  • +9
  • nianji zhan,
  • Fan Yang,
  • Jingyu Ye,
  • Dan Wang,
  • Ying Yang,
  • Sheng Zhao,
  • Bin Zou,
  • Jieping Song,
  • Fang Chen,
  • Xinlin Chen,
  • Jian Guo,
  • Peiwen Chen
nianji zhan
BGI-Shenzhen

Corresponding Author:[email protected]

Author Profile
Jingyu Ye
BGI-Shenzhen
Author Profile
Dan Wang
Hubei Maternal and child Healthcare Hospital
Author Profile
Ying Yang
BGI-Shenzhen
Author Profile
Sheng Zhao
Hubei Maternal and child Healthcare Hospital
Author Profile
Bin Zou
CNGB
Author Profile
Jieping Song
Hubei Maternal and child Healthcare Hospital
Author Profile
Fang Chen
BGI-Shenzhen
Author Profile
Xinlin Chen
Author Profile
Jian Guo
BGI-Shenzhen
Author Profile
Peiwen Chen
Author Profile

Abstract

Objective This paper proposed the method of real-time detection of CNS diseases using object recognition network that mainly detects abnormal planes in video and evaluates the performance and feasibility of the object recognition network in classifying disease planes. Design Central nervous system cases, random sampling. Setting Prenatal ultrasound images from Maternal and Child Healthcare Hospital, Hubei. Sample A total of 515 fetal with First-trimesters. Methods Compare the three different models was training by the same dataset, including Exencephaly plane, Holoprosencephaly plane, and two normal planes. Main Outcome Measures Compare the F1 scores of other classification networks on the original dataset and the ROI dataset and test the detection speed and accuracy in the real-time video. Results The our model achieved 92% accuracy in the test set, this result is higher than other models in the classification accuracy of the original data and ROI data is 56% and 87%, and can achieve real-time detection and location that to detect the speed of each frame in 0.04 seconds. Conclusions The aim is to detect disease planes of the CNS in real-time. But the model still has deficiencies and lacks confidence in the detection of certain disease levels, when there is the fake shadow in the disease plane, the model can easily detect erroneous results. This is unavoidable to small data sets, and the model also needs to continuously increase non-disease data to reduce the error rate. The results of this article have greatly increased our confidence and are instructive for future work.