Introduction

Preterm birth is a leading worldwide health problem [1, 2]. For example, in the USA, more than 1.5% of live births are of neonates who weigh less than 1500 g. Preterm neonates are more likely to have poor neurodevelopmental outcomes such as cerebral palsy and other cognitive or behavioral impairments [3,4,5,6,7].

Previous studies have shown that intervention services have significant and positive effects on the cognitive development of high-risk preterm neonates [8]. Therefore, the prediction of neurodevelopmental outcomes is crucial for clinicians to identify those neonates who would benefit from developmental interventions [8]. Many neuroimaging studies have tried to correlate imaging markers and neurodevelopmental outcomes in preterm neonates [9,10,11,12,13,14,15]. In particular, gross and microstructural characteristics that can be observed on brain MR images have been actively utilized in earlier studies. Gross abnormalities in the white matter (WM) and gray matter (GM) structures of preterm neonates were found to predict their neurodevelopmental outcome at 2 years of age [9]. Abnormalities in WM volume and microstructural signals showed high predictive value for certain neurodevelopmental outcomes [14]. Still, mild to significant psychomotor outcomes were observed in up to 16% of cases which did not exhibit visible WM injury on MR images [16].

Radiomics has been increasingly used for image-based diagnosis and prognosis, enabling the extraction and analysis of quantitative features on medical images [17]. Many radiomics studies have helped radiologists discover hidden gems of information that could aid the diagnosis, treatment, and prognosis prediction of brain tumors. While radiomics is infrequently applied to fields other than oncology in the neuroradiology field, recent studies showed that radiomics can potentially help classify ischemic group in infants [18], attention deficit hyperactivity disorder [19], or neurofibromatosis 1 subtypes [20].

We hypothesized that a radiomics analysis of WM in preterm neonates could help predict poor neurodevelopmental outcomes even without visible signal abnormalities. Thus, the objective of this study was to develop a model for predicting poor neurodevelopmental outcomes and validate its performance in preterm neonates with minimal or no visible WM abnormalities.

Materials and methods

Study participants

Study patients were selected from a prospective study performed at our institution, which was designed to analyze MR image data acquired from preterm neonates to identify imaging biomarkers that could predict neurodevelopmental outcomes. Our university institutional review board approved the study design and written informed consent was obtained from all patient guardians. All data were anonymized before the MR data were processed.

From September 2017 to December 2018, 60 neonates were enrolled. The inclusion criteria for enrollment were as follows. Preterm neonates delivered (a) at less than 28 weeks of gestational age, (b) with birth weight less than 1500 g, or (c) at less than 37 weeks of gestational age with suspected ischemic brain injury. The exclusion criteria were (a) neonates who had congenital malformation, infection, or metabolic diseases that would affect later development, (b) neonates who showed intracranial hemorrhage that was grade 3 or higher on ultrasound, or (c) neonates with any structural abnormality on a prior brain imaging study. All the enrolled preterm neonates underwent brain MRI before discharge, near or at term-equivalent age. Among enrolled 60 preterm neonates, 56 neonates successfully undergone neurodevelopmental assessment at a corrected age of 12 months. Ten neonates were excluded before radiomics analysis. Therefore, a total of 46 neonates were included for the analysis (Fig. 1).

Fig. 1
figure 1

Flowchart shows study population and exclusion criteria

Full size image

Neurodevelopmental outcomes and clinical variables

One examiner evaluated the neurodevelopmental abilities of neonates using the Korean-Bayley Scales of Infant and Toddler Development, second edition (Bayley II) when each infant reached a corrected age of 12 months. Bayley II includes the psychomotor development of infants. The Psychomotor Development Index (PDI) assesses gross and fine motor skills. A delay in development was defined by a score that was more than 1 standard deviation below the normative mean (PDI score < 85). Clinical variables including birth history were collected from the electronic medical records. Questionnaires were complete by parents for the assessment of maternal education and socioeconomic level as those were associated with neurodevelopmental outcomes of preterm neonates [21, 22].

MRI protocols, regional segmentation, and radiologic assessment

Brain MRI was performed using a 3-T scanner (GE MRI 750w, GE Healthcare, Milwaukee, WI). The MRI examination included a 3D T1-weighted fast spoiled gradient-echo sequence, multiparametric quantitative MRI based on the multi-echo, multi-delay saturation recovery spin echo, and DTI. For the radiomics studies, we used synthetic T1-weighted images (T1WI) and T2-weighted images (T2WI) were generated from the multi-echo, multi-delay data. The imaging protocol is further described in Appendix 1. We obtained regional volumes (total brain, cortical GM, deep GM, and WM) using infant FreeSurfer [23, 24]. Regions of interests were drawn in both of the posterior limbs of internal capsules (PLICs) on a fractional anisotropy (FA) map derived from DTI [25, 26]. All scan results were evaluated by two radiologists, a pediatric neuroradiologist with 10 years of experience and a neuroradiologist with 13 years of experience. Both radiologists were unaware of the neonates’ perinatal history and ultrasonographic findings. White-matter abnormality was rated and classified into two categories: “none or mild abnormality” or “moderate to severe abnormality” [9]. The final classification was made in consensus.

Segmentation, feature extraction, and selection for radiomics analysis

The overall pipeline of the radiomics analysis is illustrated in Fig. 2. Segmentation was performed semi-automatically on the WM region by a radiologist. Radiomics features were defined according to the Pyradiomics Python package, version 3.0 [27]. A total of 11 image types were derived from the standard normalized original T1WI and T2WI and their filtered images. Filtering was done with wavelet-transformed (eight wavelet decompositions) and Laplacian of Gaussian-filtered (LoG) (two values of sigma) images. Radiomics features were extracted from both T1WI and T2WI, resulting in a total of 1906 features. Segmentation and feature extraction methods are described in detail in Appendix 2.

Fig. 2
figure 2

The overall pipeline of the radiomics analysis. Part 1 includes image acquisition and segmentation. Masks for regions of interest for white matter were acquired based on T2WI using the semi-automated method. Part 2 includes image normalization and filtering. A total of 11 image types (2 Laplacian of Gaussian-filtered images, 8 decompositions of wavelet-transformed images, and the original image) which were all derived from normalized T1WI and T2WI. Part 3 includes extraction of radiomics features. All of the derived image types were used to extract first-order and second-order features. The shape features were extracted from the binary white matter mask. Part 4 includes feature selection based on the minimum redundancy maximum relevance algorithm and the prediction performance of the model was evaluated by ROC analysis with the selected radiomics features

Full size image

The minimum redundancy maximum relevance (MRMR) algorithm was used to select features of high relevance to train the prediction model. Specifically, the top 20 relevant features were used to build prediction models based on T1WI and T2WI, respectively. For another model based on a combination of T1WI and T2WI, the top 10 relevant features of each T1WI and T2WI were used.

Model construction and validation

A tree ensemble classifier was used as a prediction model for the binary classification of PDI. The random undersampling boost (RUSBoost) algorithm was used for ensemble aggregation due to its capability to handle imbalanced data with discrete class labels like the training data used in this study [28]. The prediction model was trained with radiomics features selected with the MRMR algorithm in 46 data sets. Eightfold cross-validation was done to evaluate the generalized performance of the prediction model. The receiver operating characteristic (ROC) curve analysis was used to explain the prediction performance of the selected radiomics features. The optimal area under the curve (AUC) was computed for each of the three cases (T1WI, T2WI, and combination of T1WI and T2WI). Clinical or imaging variables with statistical significance on multivariate analysis were additionally included and the prediction performance was computed.

Statistical analysis

Continuous variables were found to be nonparametric. Clinical and imaging characteristics of the neonates with and without delayed psychomotor development were compared using the Mann-Whitney U test. To find significant clinical and imaging variables, we performed univariate and multivariate analyses. Results were considered to be statistically significant if p < .05. SPSS version 25.0 (IBM) was used for analysis.

Results

Clinical and imaging characteristics of neonates

A total of 46 neonates (median gestational age, 29 weeks; 26 males) were included. The clinical and imaging characteristics of the neonates are summarized in Table 1. Of the 46 neonates, 13 (28.3%) showed delayed psychomotor development and 33 (71.7%) showed normal psychomotor development. Birth weight (p = .002), gestational age (p = .001), corrected gestational age at MRI scan (p = .049), regional brain volumes (p = .006–.021), and FA value in the left PLIC (p = .035) were significantly different between the neonates with and without abnormalities. Conversely, gender (p = .749), maternal education level (p = .497), and FA value in the right PLIC (p = .262) were not significantly different between the neonates with and without delayed psychomotor development. The imaging characteristics of the neonates assessed by the two radiologists did not significantly differ between the two groups (p = .283). Among the neonates with delayed psychomotor development, there was one neonate (1/13) with moderate to severe WM abnormality. Among the neonates with normal psychomotor development, there was no neonate with moderate to severe WM abnormality.

Table 1 Clinical and imaging characteristics of the preterm neonates
Full size table

Feature selection

Table 2 shows the top 10 relevant image features extracted from T1WI and T2WI. Small dependence low gray level emphasis (gray level dependence matrix [GLDM]) obtained from the wavelet transformed with high-low-high-pass filter T1WI and cluster shade (gray level co-occurrence matrix [GLCM]) obtained from the LoG T2WI ranked as the most relevant feature in each case. In T1WI, seven features were obtained from the wavelet-transformed images, two features from the LoG images and one feature from the original images. In T2WI, one feature was obtained from the original images, one feature from the LoG images and eight features from the wavelet-transformed images. Figure 3 demonstrates the heatmap of the top 10 features selected from T1WI and T2WI, respectively, arranged by the binary class of PDI.

Table 2 List of significant MR radiomics features to classify psychomotor development in preterm neonates
Full size table
Fig. 3
figure 3

Heatmap of the top 10 radiomics features of each T1WI and T2WI. Each column corresponds to radiomics features sorted by rank, and each row corresponds to the patients sorted by binary class. (PDI < 85: delayed psychomotor development, PDI ≥ 85: normal psychomotor development)

Full size image

Model performance

Figure 4a shows the ROC curves and AUC values for the prediction of poor outcomes in preterm neonates based on T1WI, T2WI, and combination of T1WI and T2WI. The prediction model performed best with T1WI, providing an AUC of 0.925 (95% confidence interval [CI]: 0.85–1.0), sensitivity of 84.6%, specificity of 78.8%, and accuracy of 80.4%. The prediction model based on T2WI resulted in an AUC of 0.834 (95% CI: 0.72–0.95), sensitivity of 76.9%, specificity of 75.6%, and accuracy of 76.1%. The prediction model based on the combination of T1WI and T2WI resulted in an AUC of 0.902 (95% CI: 0.82–0.99), sensitivity of 92.3%, specificity of 81.8%, and accuracy of 84.8%. The confusion matrix evaluated by each prediction model is shown in Supplementary Figure S1.

Fig. 4
figure 4

The receiver operating characteristic curves and areas under the curve for three cases (a) T1WI, T2WI, and combination of T1WI and T2WI, and for four cases (b) T1WI, T1WI with gestational age, T1WI with the DTI parameter, and T1WI with both variables. GA, gestational age

Full size image

Results from multivariate and univariate analyses using clinical and imaging variables are shown in the Supplementary Tables. The significant variable on multivariate analysis was gestational age. Regional brain volumes were not a significant factor predicting poor psychomotor outcomes. Higher FA value in left PLIC resulted statistically insignificant but high odds ratio (134.5). We added gestational age, FA value in the left PLIC, or both in the prediction model based on T1WI. Adding gestational age or FA values did not result in higher prediction performance (Fig. 4b). The prediction model based on T1WI with gestational age resulted in an AUC of 0.900 (95% CI: 0.81–0.99), sensitivity of 84.6%, specificity of 78.8%, and accuracy of 80.4%. The model based on T1WI with FA value in the left PLIC resulted in an AUC of 0.908 (95% CI: 0.82–0.99), sensitivity of 84.6%, specificity of 75.8%, and accuracy of 78.3%. The prediction model based on T1WI including both gestational age and FA value in the left PLIC resulted in an AUC of 0.897 (95% CI: 0.81–0.99), sensitivity of 84.6%, specificity of 78.8%, and accuracy of 80.4%. The confusion matrix evaluated by each prediction model is shown in Supplementary Figure S2.

Discussion

In this study, we demonstrated that a radiomics analysis of the WM of preterm neonates on MRI enables the prediction of poor neurodevelopmental outcomes. This was achieved through automated extraction of important voxel-wise imaging features and training of a prediction model for binary classification. We also found that T1WI is more effective than T2WI for predicting poor neurodevelopment when data is fed into the developed classifier.

It should be noted that our radiomics analysis was performed on preterm neonates and most of the neonates did not present with moderate to severe WM injury (45 out of 46). The radiomics analysis was able to successfully predict the neurodevelopmental outcomes of these neonates. The rate of neurodevelopmental impairment among those with no to mild WM injury in our study (26.7%) was comparable to the rate reported in a previous study (15–27%) [29]. This suggests that although moderate to severe WM injury is a good predictive imaging marker for poor neurodevelopmental outcomes, about one-fourth of preterm neonates without significant WM injury will still eventually have problems with their neurodevelopment. This finding underscores the importance of predicting neurodevelopmental outcomes in neonates without significant WM injury.

Previous studies which assessed WM to predict neurodevelopmental impairment in preterm neonates focused on abnormal signal intensity or volume loss [14, 15, 30]. The sensitivity of our model using T1WI (84.6%) was higher than that of previous studies (38–50%) using visual assessments for WM injury, but the specificity of our model was lower (78.8% vs 82–100%) [9, 15]. It should be noted that we included only one case with significant WM injury per visual assessment, which makes it difficult to make direct comparisons with previous studies that included various degrees of WM injuries. Nonetheless, the higher performance of our prediction model in the absence of significant WM injury may be attributed to the highly quantified WM features which is a trait not possible with the conventional visual scoring system [9]. We believe that differences in the extracted radiomics features between normal and poor neurodevelopmental outcomes arose mainly from the microstructural changes caused by preterm birth.

Including gestational age or DTI parameters did not have additional benefits for predicting psychomotor outcomes in this study, emphasizing the value of radiomics feature analysis. A recent study showed that clinical parameters such as gestational age and social economic status had a higher value than imaging parameters of brain volume when predicting cognitive outcomes [31]. On the other hand, the same study showed brain volume as a significant factor when predicting psychomotor outcomes [31]. Although our assessment of regional volumes showed significant differences between the normal and poor psychomotor outcome groups, these parameters were insignificant on multivariate analysis. Similarly, although a DTI parameter was significantly different between normal and poor psychomotor outcome groups as in previous studies [25, 26], it was insignificant on multivariate analysis. The highest AUC for predicting poor psychomotor outcomes using regional brain volume or clinical variables in the past study was 0.64 [31]. We can postulate that radiomics features derived from WM may have higher predictive values than regional brain volume, DTI parameters, or clinical variables such as gestational age.

In our study, the significant radiomics features of T1WI and T2WI turned out to be GLDM, GLCM, and gray level size zone matrix (GLSZM). GLDM is a texture feature which quantifies gray level dependency that is defined as the number of connected voxels within a certain distance from the center voxel in an image [32]. GLCM is also a texture feature which reflects signal heterogeneity by calculating the joint probability distributions of pixel pairs [32]. Our results which showed significant difference in the GLCM of WM according to neurodevelopmental outcomes were in line with previous studies that showed the predictive value of WM heterogeneity in adults [33,34,35]. GLSZM represents the number of connected voxels with the same intensity values [32]. Many major features were significantly more relevant when obtained from wavelet-transformed images, especially from high-low-high (HLH) decomposition. We believe that texture analysis will show better performance when various features are combined.

There are some limitations to this study. First, the data set was of relatively small size (n = 46). Further validation with a larger data set is needed before clinical application is possible. Second, we only validated the model performance internally without external validation. Related future work should include test data from different institutions for validation and optimization of image acquisition and analysis protocols. Third, the use of synthetic T1WI and T2WI instead of conventional T1WI or T2WI may limit the transferability of our model to other institutions. Although synthetic and true contrast images are thought to be similar in children [36], the exact signal intensity and distribution of images are different and this would affect the radiomics features. Therefore, signal intensity needs to be normalized to a fine degree during the entire procedure to validate our findings with conventional T1WI and T2WI. Finally, we did not subcategorize the study subjects according to WM injury before analysis. The subjects were prospectively enrolled and those with high-grade intracranial hemorrhages were excluded from the study population. Still, there was one subject with moderate to high-grade WM injury. This suggests that although few, some subjects with different degrees of WM injuries might be included in the radiomics analysis which may affect the radiomics features.

In conclusion, this study showed that radiomics-driven analysis of early brain MR images can predict poor psychomotor outcomes in preterm neonates. Our prediction model might be of potential use in quantitative analysis using T1WI and T2WI and in the absence of significant WM abnormalities.