Radiomics based of deep medullary veins on susceptibility-weighted imaging in infants: predicting the severity of brain injury of neonates with perinatal asphyxia

A patient with hypoxic-ischemic encephalopathy (HIE) may suffer from neurological problems as a result of neonatal asphyxia [1]. Globally, it causes 1 to 8 neonatal deaths per 1000 live births, making it one of the leading causes of neonatal morbidity and mortality [2]. Brain injuries caused by HIE lead to a serious condition, rapid progress and poor prognosis. Some patients can experience varying degrees of neurological sequelae. Sarnat scores are commonly used to evaluate neonatal HIE severity. Based on the clinical signs and electroencephalograms, according to Sarnat criteria, HIE can be mild (stage 1), moderate (stage 2), or severe (stage 3) [3]. HIE can be more effectively treated and predicted over the long term if the scope and severity of the disease are clarified and the diagnosis staged [4].

Although the Sarnat criteria was used to evaluate neonatal asphyxia clinically, the Sarnat score is subjective and easy to be impacted and affected by non-asphyxia factors. HIE is most commonly imaged with magnetic resonance imaging (MRI). There are, however, variations in the imaging patterns of HIE injury depending on the severity, the gestational age, the duration of the injury, and when the imaging is performed. Furthermore, in conventional T1WI and T2WI sequences, infants with unmyelinated brain structures and high-water content have difficulties discriminating pathologic signals and evaluating brain MRIs. In clinical practice, susceptibility-weighted imaging (SWI) is increasingly applied to HIE. The SWI is a high-resolution, 3D gradient echo imaging sequence, which uses the difference of magnetic sensitivity between tissues to display the paramagnetic properties of blood products through phase post-processing. Therefore, it is more sensitive to the detection of intravenous deoxyhemoglobin and extravascular blood products. HIE can cause contracture of small blood vessels in the brain and increase vascular permeability. Many factors cause vascular rupture and bleeding. Hypoxia and ischemia of brain tissue can cause congestion of cerebral veins and increase venous pressure, which is easy to cause dilation of deep cerebral veins in varying degrees [5]. It has been reported that DMV prominence is associated with both sinovenous thrombosis and white matter lesions (WM) in infants [6]. There have been few objective methods used to quantify DMV, even though it is thought to be a pathologic change observed during SWI.

MRI is an indispensable imaging modality for various clinical conditions, but it may only provide limited information due to human perception limitations. Radiomics is the extraction, analysis, and mining of medical data from digital medical images using high-throughput computational methods [7]. By selecting features from these data, biomarkers can be constructed for disease prediction and diagnosis. It is possible to obtain potentially valuable information through radiomics beyond the limitations of human analysis [8]. Numerous studies have investigated radiomics' clinical applicability. In various cancers, radiomics features have been shown to be useful imaging predictors of diagnosis, treatment response, prediction, and prognosis [9‐11], and some studies use texture analysis to evaluating ischemic changes in neonates [12, 13]. There has been no previous data assessing Rad-score and combined nomogram model for HIE in neonates predicting the severity of brain injury. In this study, we developed and validated a combined nomogram model that extracted radiomics features from SWI to determine the severity of brain injury in neonates with HIE.

Neonatal HIE is a type of brain injury caused by cerebral blood supply and gas exchange disorders in perinates. In the severe stage, it can lead to irreversible brain injury or even death. The main changes included neurocyte degeneration, necrosis, brain edema, cerebral infarction, and intracranial hemorrhage. Ischemia and hypoxia can cause cerebral vasospasm, increased vascular permeability, and brittleness, and a variety of factors can cause vascular ruptures and hemorrhages [15]. Cerebral hypoxia and ischemia lead to cerebral venous congestion and increased venous pressure, which is easy to cause the expansion of deep cerebral veins and cortical veins in varying degrees [16]. In both healthy and pathologic subjects, SWI can better delineate cerebral venous structures compared to conventional MRI scanning. DMVs drain blood from WM to subependymal veins in the cerebral venous system [17]. Therefore, it is easier to display DMVs area on SWI and obtain ROI more accurately. In this study, we used manual segmentation to extract and select the image radiomics features with predictive values by outlining the ROI of neonatal SWI images.

Using a radiomics model derived from SWI features and deep medullary veins, we demonstrated that moderate-to-severe neonatal HIE could be accurately diagnosed. We observed 15 potential predictors imaging features that were significantly related to it, including wavelet_HLH_firstorder_Median, wavelet_LHH_gldm_SmallDependenceLowGrayLevelEmphasis, and wacelet_LHH. These features have the highest predictive value. A radiomics model was developed and validated for the accurate diagnosis of moderate-to-severe HIE in neonates. In the validation cohort, the AUC of radiomics signatures was 0.80 (95% CI 0.540–0.797), accuracy was 0.687, sensitivity was 0.620, and specificity was 0.714. In the training cohort, the AUC of radiomics signatures was 0.84 (95% CI 0.704–0.850), accuracy was 0.784, sensitivity was 0.784, and specificity was 0.783. As a result, the radiomics model can predict moderate-to-severe neonatal HIE with a certain degree of accuracy. As well, urea nitrogen was the only independent factor that could distinguish mild and moderate-to-severe neonatal HIE in our study. To diagnose moderate-to-severe neonatal HIE accurately, we developed and validated a combined nomogram model. Both the training and validation cohorts of the combined nomogram model did not achieve a higher AUC value than the radiomics model, suggesting the clinical features did not significantly improve the predictive value of radiomics model in distinguishing mild and moderate-to-severe neonatal HIE. Based on DCA results, the combined nomogram and radiomics model performed better than the clinical model for moderate-to-severe neonatal HIE prediction. Additionally, the training cohort and validation cohort calibration curves were constructed. According to the calibration curves based on the validation cohort, the proposed models have a favorable classification performance.

Some studies used MRI score ability to detect HIE abnormalities, although conventional and further MRI techniques have described HIE features in neonates with HIE [18‐20]; our study is the first article that uses a nomogram model to distinguish mild HIE form moderate-to-severe HIE. In the past, there were few studies on the use of radiomics to analysis neonatal HIE. In previous studies, infants with ischemic injuries can be differentiated based on texture features. Weiss et al. are working on combining radiomics with machine learning to detect lesions and predict outcomes in patients with HIE; however, they have not published their findings at the time of writing [21]. Kim et al. [12] only used histogram analysis to study the feasibility in infants with ischemic injury, and their study only used a limited number of infants. Their AUC was 0.865, which was similar to our study. Based on the texture of the basal ganglia and thalami in neonates, Fatma et al. [13] demonstrated accurate diagnosis of moderate-to-severe HIE, with a sensitivity of 95% and accuracy of 94.3%. However, it is difficult to directly compare the predictive power of our model to Fatma et al. models. In this model, previous studies only used texture features. Texture features are very useful for identifying target images with obvious texture features; however, its main disadvantage is that when the resolution of the image and the illumination of the target change, the texture of the target image may produce a large deviation and affect the classification effect. A larger sample size of 170 patients was used in our study, compared to a sample size ranging from 7 to 35 patients in previous studies. A small sample size will lead to overfitting and affect the authenticity of the data. In our study, we first added clinical independent factor combined radiomics features to establish a nomogram model. The nomogram model visualizes the radiomics features and clinical predictors and provides a simple and easy-to-use tool for the individualized prediction of HIE in neonates and predict the severity of HIE.

Despite a lack of understanding about the underlying mechanism for radiomics and nomograms that reflect HIE stages, we speculate that radiomics and nomogram can reveal the micro-changes in hypoxic injury. A wide variety of pediatric neurological disorders have been studied using SWI in previous studies, including chronic ischemia, developmental venous anomalies, microhemorrhages, convulsive disorder, and hypoxic-ischemic injury [5, 22, 23]. For ischemic changes, SWI can clearly show the abnormal dilation of small veins. In the early stage of HIE, brain histopathology shows cerebral edema and small vein hyperemia and dilation. After cerebral hypoxia, small artery reactive dilation and low brain oxygen uptake rate resulted in hemodynamic and blood compensatory damage at the blood level of vascular tissue. This led to an increase in the proportion of small vein deoxyhemoglobin, which was shown as small vein dilation on SWI [24, 25]. Kitamura et al. have suggested that the degree of deep medullary vein dilatation in children with hypoxic-ischemic brain injuries can indicate the prognosis of the nervous system.

There are several limitations to this study. First, it is a retrospective study done by a single center with no external validation; there can also be a limited generalizability due to case selection bias. Second, although this study included a relatively large number of neonates, the cohort was still small when compared with other radiomics studies, especially the mild HIE group; our results may not be generalizable due to these factors. To validate our findings, we need a large-scale, prospective, multicenter study. Thirdly, as we know, in neonates with perinatal asphyxia, the combination of T1WI, T2WI, and DWI was the most commonly used method to detect cerebral damage [26]. To enhance the clinical impact of these models, we need to investigate whether they can further improve diagnostic efficiency of HIE severity, and whether they can predict neurodevelopment outcomes in HIE along with T1WI, T2WI, and DWI. Fourth, manual segmentation was used to delineate the ROI of the DMVs, but automatic or semi-automatic segmentation was not used for comparison and verification, which had a certain subjective impact. Finally, a severe perinatal asphyxia also affects the cerebral cortex, basal ganglia, and thalami [27], and was not considered in this study because the discrimination of the ROI in those areas would not be reliable on SWI. The above deficiencies need to be further improved by further research.

For the classification of mild HIE and moderate-to-severe HIE in neonates, the combined nomogram model and radiomics model can be reliable and effective. Even if there is no visually detectable difference between the DMVs on SWI, they may generate objective features which can indicate differences. In our study, we suggest that radiomics analysis of SWI can be a useful tool in predicting the severity of brain injury of infants with HIE.