Susceptibility weighted image analysis methods for hypoxic-ischaemic encephalopathy prognosis

Abstract

Neonatal hypoxic-ischaemic encephalopathy (HIE) is a major cause of newborn deaths and neurodevelopmental abnormalities around the world. Susceptibility weighted imaging can provide assistance in the prognosis of neonatal HIE. The propose of this research is to develop a new automated system to assess neonatal brain injury and developmental outcome by detecting and analysing vessel features in SWI images. In this research, a dataset of SWI images acquired from 42 infants with neonatal HIE is used for feature extraction. Firstly, the ridges representing the veins in the SWI images are detecting to obtain features including the width, intensity value, length of veins, and Hessian eigenvalues for ridges. The normalized histograms of these features are used as feature vector for classification. Individual or concatenated feature vectors are fed into kNN and random forest classifiers to predict the neurological outcomes of infants with HIE at the age of 24 months. We select the balanced SWI dataset to avoid the bias. The feature vectors containing width, intensity, length and eigenvalue show a promising classification accuracy of 78.67% ± 2.58%. Then we use the feature vectors to train support vector regression and random forest regression models for predicting the motor score and cognitive score of infants with HIE assessed by Barley-III at the age of 24 months. Our mean relative errors for cognitive and motor outcome scores are 0.113±0.13 and 0.109±0.067 respectively. The features derived from the ridges of the veins are good predictors of neurological outcome in infants with neonatal HIE. Further, we design a supervised classifier for automatic prognosis of automated detection of SWI signs of HIE. This classifier also enables to determination of brain regions which have been affected by hypoxicischaemic by extracting appropriate features from SWI images. Our classifier can classify the veins in the SWI images into normal and abnormal group by clinical assessment outcomes. The number and location of abnormal veins in the brain of HIE neonates will predict the neurodevelopmental outcomes of infants with HIE at the age of 24 months. Our classifier proposed in this study demonstrates a superior performance in HIE prognosis for the dataset associated with cognitive and motor outcomes. The accuracy of early prediction of motor outcome at 2 years of age using SWI images in newborns by our classifier achieves 75% ± 13.9%. We also employed the linear regression, polynomial regression, and support vector regression model to predict outcomes and the lower mean relative errors for motor and cognitive outcomes are 0.088±0.073 and 0.101±0.11 respectively. Then we extract the feature vectors of global and local brain by the histogram of oriented gradient descriptor. We obtain the brain regions associated with motor and cognitive function by image registration. The histogram of oriented gradient feature vectors of these brain regions are fed into the kNN and random forest classifiers to predict the motor and cognitive outcome. The result shows an effective classification for cognitive outcome, which achieved an accuracy of 76.25±10.9. In addition, we propose a convolutional neural network model to classify the SWI images with HIE. Due to the lack of a large dataset, transfer learning method with fine-tuning a pre-trained ResNet 50 is introduced. we train a convolutional neural network model to classify the SWI images with HIE. Due to the lack of data, transfer learning method with fine-tuning a pre-trained ResNet 50 network is introduced. The balanced datasets are selected randomly to avoid bias in classification. Then we develop a rule-based system to improve the classification performance, with an accuracy of 0.933 ± 0.086. We also compute heatmaps produced by the Grad-CAM technique to analyze which areas of SWI images contributed more to the classification results. Our research demonstrates that the features derived from the vascular ridges improve the prognostic value of SWI images in HIE. Furthermore, our findings suggest that it is possible to predict neurological, motor, and cognitive outcomes by numerical analysis of their neonatal SWI images and to identify brain regions on SWI affected by HIE.

More information

Identifiers

ORCID for Mark Nixon: orcid.org/0000-0002-9174-5934

Catalogue record

Export record