The role of MRI-based texture analysis to predict the severity of brain injury in neonates with perinatal asphyxia

Introduction

Hypoxic-ischemic encephalopathy (HIE) is one of the most devastating results of perinatal asphyxia.¹ Sarnat score is most commonly used clinical scoring system for evaluating the severity of neonatal HIE. Patients are classified as mild (Stage 1), moderate (Stage 2), and severe (Stage 3) HIE according to the Sarnat criteria, which is based on the clinical signs and electroencephalographic results.² Determining the severity of brain injury and staging of HIE is crucial, as it plays an important role in establishing neuroprotective therapeutic options and predicting long-term outcome.³

The imaging patterns of hypoxic injury vary depending on the severity and duration of insult, gestational age, and the timing of imaging. MRI is the best modality for evaluating suspected HIE.⁴ Abnormal signal intensities can be usually detected in basal ganglia and thalamus, which are known as the vulnerable sites to oxidative stress, in neonates with severe HIE.^4,5 However, it is reported that neurologic deficits may also occur in some neonates with normal brain MRI findings.^6,7

Although MRI is an indispensable imaging modality for various clinical situations, it can provide limited information that is based on human eye perception. Furthermore, unmyelinated brain structure with high-water content in newborns makes the discrimination of the pathologic signal and the evaluation of brain MRI more difficult. Recently, texture analysis (TA) as a radiomic approach has been widely used for non-invasive quantitative evaluation of different pathological conditions. TA allows extracting the data of gray-level intensity, the position of pixels, and the arrangement and interrelation among voxel intensities from medical images.^8,9 The radiomic features are generally grouped into as conventional, first-order, and second-order features. Since TA has the ability to expose microstructural changes that cannot be noticed by the naked eye, there have been numerous studies that investigated the clinical applicability of it.^10–14

There has been no previous data regarding the TA in neonates with perinatal asphyxia. In this study, we focused on the neonates with HIE, and investigated the potential role of MRI-based TA in predicting the severity of brain injury.

Methods

Image interpretation

DWI, T₁W, and T₂W images were retrospectively evaluated by two radiologists (FCS is a 3-year-experienced pediatric radiologist and OS is a 9-year-experienced radiologist) blinded to clinical data. In cases of the disagreement between the radiologists, a 21-year-experienced pediatric radiologist (HG) evaluated the images. The readers searched for the findings indicating the severe perinatal asphyxia such as a restriction of diffusion in the basal ganglia, thalamus, or the posterior limb of the internal capsule (PLIC) on DWI,^5,16 increasing signal intensity of the basal ganglia and thalamus on T₁W images, and an absence of hyperintensity of the PLIC on T₁W images (absent or present)^5,17 (Figure 1). The findings were determined as present if the signal intensity abnormality was observed either unilateral or bilateral. The criteria for image interpretation were presented extendedly in Supplementary Material 1.

Download Figure

Texture analysis

Texture features was calculated on two-dimensional sectional images using LifeX software (www.lifexsoft.org).¹⁸ The texture features of the basal ganglia and thalami were investigated. Two radiologists (FCS and OS) performed the measurements, independently. Axial plane images on T₁W, T₂W sequences, and ADC maps were exported in Digital Imaging and Communications in Medicine format from the medical database to LifeX for region of interest (ROI) delineation. The uniformity was provided by adjusting the gray-levels to 128 (7 bits). Intensity rescaling values were conducted automatically between mean ± 3 standard deviations (SDs). Pixels with greater or lesser values than mean ± 3 SDs, in the beginning, were set to mean ± 3 SDs. Lastly, the voxel values in three directions were set as X, 0.7 mm; Y, 0.5 mm; Z, 1 mm after calculating their mean ± 3 SDs. The ROIs for TA of the basal ganglia and thalamus were drawn manually and bilaterally in each sequence on axial two-dimensional images (Figure 2). 47 TA features were measured and calculated for basal ganglia and thalamus, separately. The features were:

• Seven conventional
• Two shape
• Six histogram (HISTO),
• Seven gray-level co-occurrence matrix (GLCM),
• 11 gray-level run-length matrix (GLRLM),
• Three neighborhood gray-level different matrix (NGLDM),
• 11 gray-level zone length matrix (GLZLM).

Download Figure

All the process was taken approximately 10 min per patient. Further details about the features are presented in Supplementary Material 1.

Results

Logistic regression analysis

Binary logistic regression analysis was used to specify the independent predictors for the moderate-to-severe HIE among all texture features that had p values lesser than 0.001. Using a basal ganglia HISTO_entropy log-10 value of higher than 1.8 or a HISTO_entropy log-2 value of higher than 5.98 on the ADC maps were found as independent features for predicting the moderate-to-severe HIE (p < 0.001; odds ratio [OR], 266; confidence interval [CI], 15.2–4628.4). The absence of hyperintensity of the PLIC on T₁W images (p = 0.012; OR, 17.11; CI, 1.8–156.2) was found as an independent finding among the radiological features.

Table 5 and Figure 3 summarize the comparison of the diagnostic performances of the independent predictors and their combination. Using only a basal ganglia HISTO_entropy log-10 value of higher than 1.8 or a HISTO_entropy log-2 value of higher than 5.98 on the ADC maps had equally diagnostic yields when compared to its' combination with the absence of hyperintensity of the PLIC on T₁W images.

Download Figure

Table 5. The diagnostic performances of the texture features with the cut-off values, radiological findings, and their combined-model to predict the moderate-to-severe HIE^a

	Sensitivity (%)	Specificity (%)	PPV (%)	NPV (%)	Accuracy (%)
ADC map-HISTO_Entropy-log10b	95 (73.0–99.7)	93.3 (66.0–99.6)	95	93.3	94.3
Absence of hyperintensity of the PLIC on T1 weighted images	55 (32.0–76.1)	93.3 (66.0–99.6)	91.6	60.8	71.4
The combined-modelc	95 (73–99.7)	93.3 (66–99.6)	95	93.3	94.3

ADC, apparent diffusion coefficient;HIE, hypoxic-ischemic encephalopathy; HISTO, histogram; PLIC, the posterior limb of the internal capsule.

aNumbers in parentheses are 95% confidence intervals.

bHISTO_Entropy-log10 cut-off value on the ADC maps was obtained from the basal ganglia. The cut-off value was higher than 1.8. Using a HISTO_Entropy-log2 cut- off value (>5.98) on the ADC maps had same results as using a HISTO_Entropy-log10.

cThe combined-model indicates to the combination of the cut-off value of HISTO_Entropy-log10 on the ADC maps and absence of hyperintensity of the PLIC on T₁ weighted images.

Discussion

Our study demonstrated that MRI-based TA of the basal ganglia and thalami allows for accurate diagnosis of moderate-to-severe HIE in neonates. We found that using only a HISTO_entropy log-10 value as >1.8 or a HISTO_entropy log-2 value as >5.98 that were obtained from the basal ganglia on the ADC maps had an exquisite diagnostic performance with a sensitivity of 95%, a specificity of 93.3, and a diagnostic accuracy of 94.3. The combined-models that were used together with the HISTO_entropy log-10 cutoff value, either the radiological findings or other significant texture parameters, had not significant contributions for the differentiation of moderate-to-severe HIE from mild HIE when compared to using HISTO_entropy log-10 alone. We also demonstrated that several texture parameters from basal ganglia and thalami had high diagnostic accuracy for predicting the severity of HIE. Besides these, we presented the texture feature values of the basal ganglia and thalami in normal neonates with comparisons of those of HIE groups.

Although conventional and further techniques of brain MRI findings in neonates with HIE have been described,^{1,5,17,19–25} to our best knowledge, this is the first article that evaluated the diagnostic yield of TA for differentiating moderate-to-severe HIE from mild HIE. In 2019, Weiss et al²⁶ proclaimed that they will work on a study protocol including radiomics that will provide MRI‐based HIE lesion detection and outcome prediction. However, their results have not been published yet. Apart from this, Kim et al²⁷ investigated the feasibility of the deep medullary veins TA in infants with developmental and ischemic injury. The authors performed their study by only using first-order histogram analysis and including very few infants. Despite several technical and methodologic differences when compared to our study, they also reported that the texture features can potentially be used as quantitative markers for differentiating infants with ischemic injury through deep medullary vein changes.

A series of quantitative imaging features can be extracted from MRI-based TA. The first-order statistical feature, also called the histogram, is the distribution of voxel intensities within the ROI by various basic features, such as skewness, kurtosis, entropy, and energy of gray level intensity. Entropy is defined as randomness and a measure of the amount of uncertainty in a source.²⁸ Although it is not possible to interpret precisely, we can claim that a larger value of entropy may indicate a stronger heterogeneity of intensity values since higher results than threshold value point out the moderate-to-severe HIE according to our findings. In addition to the excellent diagnostic performance of the HISTO_entropy values, second-order features including GLCM, NGLDM, and GLZLM values, which describe the statistical relationships between pixels or voxels, were also found as notably useful texture parameters for discriminating between the mild and moderate-to-severe HIE in this study. Considering the diagnostic utility of the features as well as the short application time, we believe that MRI-based radiomics may be used as a diagnostic tool in daily practice.

According to our results, there were several statistically significant texture parameters between the normal neonates and mild HIE and moderate-to-severe HIE. Although, the deep gray matter injury does not an expected finding in the mild hypoxic injury,⁵ we could speculate that TA might show the imperceptible findings in the deep gray matter on conventional MRI.

Although the involvement of the basal ganglia and thalamus usually occurs in the setting of the moderate-to-severe hypoxic injury, the detection of the signal abnormalities is highly related to the timing of the imaging. There was no difference regarding to the time from birth to MRI between the mild (5.6 days) and moderate-to-severe HIE (5.9 days) groups (p = 0.419). Imai et al²⁹ investigated the early (1–3 days) and late (4–7 days) MRI changes in the thalamus and basal ganglia in neonates with perinatal asphyxia. They reported that ADC values were lower in the thalami in the early period while the values were lower in the basal ganglia in the late period. Although we found significant thalamic texture feature values in the ADC maps for differentiating the groups, the area under curve results, the significance of values, and the number of significant parameters were superior for the basal ganglia measurements. It may be related to the mean time of the imaging in our study group.

Considering the radiological findings, an absence of hyperintensity of the PLIC on T₁W images was found as an independent predictor the moderate-to-severe HIE (p = 0.012). Due to the high specificity value (93.3%) of this finding, we can comment that the high signal intensity in the PLIC on T₁W images indicates to the mild HIE in neonates with perinatal asphyxia. Although the finding was found as an independent predictor, the diagnostic accuracy (71.4%) was not substantial, especially compared with the values of TA (93.3%).

Liauw et al²⁵ reported the combination of T₁W, T₂W, and DWI as the best for detecting cerebral damage in neonates with perinatal asphyxia. T₂W images are also usually used for a confirmation of the signal intensity abnormalities which are detected on T₁W images.⁵ We used these three sequences in TA for two main reasons. First and foremost, we can gain substantial information from TA which is not detected on visual inspection. The second reason was that we did not want to overlook the contributes of the sequences since the best time for the depiction of the brain damage varies in each sequence.^4,5,20

In terms of the utility of the radiomic features, new studies have been adding to the literature continuously. There are a few studies in the literature on the use of basal ganglia TA features in different diseases.^30,31 Johns et al³⁰ found that the basal ganglia texture features significantly differ in patients with amyotrophic lateral sclerosis. In another study, it is suggested that the basal ganglia TA can be used for evaluating the blood-brain integrity in small vessel disease.³¹ This is the first study which presented the diagnostic role of the MRI-based TA to predict the severity of hypoxic injury in neonates with asphyxia. We investigated the texture features in both basal ganglia and thalami on DWI, T₁W, and T₂W images to reveal the microchanges in hypoxic injury. Also, we included the control group to demonstrate the normal texture features values and to reveal microstructural changes in neonates with asphyxia with normal-appearanced brain MRI. These are the major strengths of this study.

Our study had several limitations. First, the study had a retrospective design and the study sample was small. Second, the ROIs of the basal ganglia and thalami were selected on axial images as two-dimensional. Neonates have fewer slice numbers due to their relatively smaller field of view. We thought that it was not ideal to delineate the ROI in multiple slices especially in the ADC maps. Third, although radiological findings and metabolite values on MRI-spectroscopy could be changed after the hypothermia treatment,^32,33 we did not exclude the neonates who underwent hypothermia treatment since it was one of the routine procedures for severe perinatal asphyxia in many centers. Fourth, as it is known, various imaging protocols may potentially affect the TA. To handle this issue, we used preprocessing steps to normalize the voxel sizes and homogenize the pixel discretization. However, the variety of some of the imaging parameters that could not be normalized, such as TR and TE values of the sequences, might be one of the limitations of this work. There could be more reliable results with using the same imaging protocols for all patients. In addition to this, the timing of imaging, which is an important determinant of radiological findings in patients with HIE, may also have an effect on the TA. So, further studies with different protocols are needed to verify the reproducibility and feasibility of our results. Fifth, cerebral cortex, which is one of the affected areas in severe perinatal asphyxia,³⁴ was not considered in TA since we thought that the discrimination of the ROI in the cortical gray matter would not be reliable. Sixth, the low TR values may have negatively affected the optimization of signal-noise ratio to achieve good gray-white matter differentiation.

Conclusion

In conclusion, our results suggest that using only HISTO_entropy values of the basal ganglia on the ADC maps allows for accurate diagnosis of the moderate-to-severe HIE in neonates. TA may generate objective features that are able to indicate differences in the basal ganglia and thalami, even if there is no visually detectable difference, in neonates with perinatal asphyxia. Although implementing tools for TA is relatively simple and can be applied for many imaging techniques, further studies are needed until these methods can be used reliably in clinical practice.