Introduction

Primary brain tumors are a heterogeneous group of benign as well as malignant tumors arising from the brain parenchyma and its surrounding structures. Brain tumors are the most common solid tumors in the paediatric age group and they are a leading cause of mortality and morbidity in children worldwide [1–3], exceeded only by leukaemia [3, 4]. According to most of the studies, the three most common types of paediatric brain tumors are astrocytoma, medulloblastoma, and ependymoma [1]. Overall survival in paediatric brain tumors varies with the type and grade of the tumor. It has been found that low-grade gliomas like pilocytic astrocytoma after gross total resection have 10-year progression-free survival of over 95% and have the best survival among paediatric brain tumors [5]. Children with non-disseminated medulloblastoma have an approximate 60% to 65% likelihood of survival for 5 years; however, the survival in disseminated tumors is less favourable, approximately 35% to 40% [5]. Ependymoma after gross total resection and radiotherapy is expected to have over 75% chance to survive for 5 years without any recurrence [5]. Brainstem glioma in children has a 3-year survival rate of only 5% to 15% even after treatment [6].

Early diagnosis and treatment of paediatric brain tumors significantly improve the outcomes [7]. To accurately diagnose paediatric brain tumors on imaging, specialized radiologists with experience in neuroradiology as well as in neuro-oncology are required. This niche requirement is often difficult to meet in many primary healthcare setups worldwide. Artificial intelligence (AI) is ever evolving tool in healthcare; especially in diagnostic oncology [8–11]. If accurate enough, AI-based tools can serve patients where specific skilled radiologists are not available. Recent advances in AI have made such diagnostic tools possible. Few machine learning-based algorithms have been developed in recent times [12]. Also, the role of machine learning in survival prediction and prognostication has been publicized in the field of neuro-oncology [13, 14]. The majority of brain tumor segmentation and radio genomics classification work has been published for glioblastoma in recent literature [15–17].

The purpose of this study was to study semantic and deep learning magnetic resonance imaging (MRI) features of paediatric brain tumors and to develop a deep learning based automated segmentation (AS) tool which could segment paediatric brain tumors and predict subtypes.

Materials and methods

Radiology review

The MRI semantic features were detected by a radiologist with 12 years of experience in neuroimaging. In addition to routine sequences (T1, post-contrast T1, T2, FLAIR, and DWI), the gradient echo (GRE)/susceptibility-weighted imaging (SWI) sequences were evaluated for the detection of blooming within tumors. Isointensity was labelled when the tumor signal had similar intensity as that of grey matter, hypo intensity, and hyperintensity when the tumor signal had low and high intensity as compared to grey matter respectively. High signals on isotropic (DWI) images with corresponding low apparent diffusion coefficient (ADC) values were labelled as restricted diffusion. Enhancement was quantified as mild, moderate, and severe as compared to the entire tumor volume. The enhancement was labelled as heterogeneous or homogeneous. Training data is used to train the deep neural network, and validation data is used to check the network performance and fine-tune the network. Test data was used to assess the final accuracy of the trained model. In order to make models generalize better and create variations of the data, data augmentation was performed. For each patient, MRI data of size 240 × 240 × 155 was provided with FLAIR, T1, postcontrast T1, T2, and DWI sequences. The tumor was annotated on all slices in the post-contrast phase. Annotations for tumor identification were done as shown in Figure 1. Separate annotations were also done for enhancing and non-enhancing components of the tumor as shown in Figure 2. An annotation to identify a tumor site was shows in Figure 3. The annotations were verified by the same radiologist. The deep learning model was trained to identify the tumor, location, and then segmentation to identify the enhancing and non-enhancing component of the tumor. The deep learning-based model predicted focal T1 hyperintensity as a haemorrhage. For cyst and necrosis, the model classified cyst as structures that are T2 hyperintense and T1 hypointense and shows significant suppression (> 75%) and necrosis to structures which did not show significant suppression (< 25% or no suppression) on FLAIR. Deep learning based-model predicted features was showed in Table 2. For prediction of signal intensity on T1WI, T2WI, FLAIR, cyst, necrosis, diffusion restriction, tumor location, and enhancement, they followed similar identification criteria to an experienced radiologist.

Display full size

Figure 1. 

Figure shows annotation for tumor identification. A: anterior; P: posterior; R: right; L: left

Display full size

Figure 2. 

Figure shows annotation for enhancing (green) and non-enhancing (yellow) components of the tumor. A: anterior; P: posterior; R: right; L: left

Display full size

Figure 3. 

Figure shows annotations for training the model to identify the site of the tumor. Blue: forebrain; pink: cerebellum; aqua: brainstem; A: anterior; P: posterior; R: right; L: left

Table 2.

MR features used for deep learning-based model prediction

MR sequence	Intensity pattern	Intensity pattern	Intensity pattern
T1WI	Hyperintense	Isointense	Hypointense
T2WI	Hyperintense	Isointense	Hypointense
FLAIR	Hyperintense	Isointense	Hypointense
Haemorrhage	Present	Absent	-
DWI	Restriction	No restriction	-
Cyst	Present	Absent	-
Necrosis	Present	Absent	-
Enhancement pattern	Homogenous	Heterogeneous	-
Enhancement pattern	Mild (< 25%)	Moderate (25–75%)	Severe (> 75%)
Tumor location	Forebrain	Brainstem	Cerebellum and fourth ventricle

Display full size

-: blank cell

Development of deep learning algorithm and classification

The AI pipeline used two different models in a sequence. The first model performed lobe segmentation and the second model segmented tumors in the scan. The first model used a variation of three-dimensional (3D) U-NET [18] which takes a 256 × 256 × 24 voxel grid as input. Before feeding the data to the model, we clipped the voxel values from a range of 0 to 2,030. After this, the voxel values were standardized with a mean of 730 and a standard deviation (SD) of 361. The network gave three channels as output. These output labels were used to label each voxel of the 3D input as forebrain, brainstem, and cerebellum. The training and validation curve for the lobe segmentation network is shown in Figure 4. These predictions were later used to build the heuristic for post-processing of the results from the pipeline. The second model was a variation of 2D U-NET [18]. We processed the data at this stage in two steps. In the first step, scan slices were taken in the shape of 512 × 512 × 1 as input and predicted tumor segmentation. The second step took predictions of the first step and second channel of input along with the original slice and predicted two channels representing enhancing and non-enhancing components of the tumor. The training and validation curve for stage 1 and stage 2 network for tumor segmentation is shown in Figure 5 and Figure 6 respectively.

Display full size

Figure 4. 

Training and validation curve for lobe segmentation network

Display full size

Figure 5. 

Training and validation curve for stage 1 network for tumor segmentation

Display full size

Figure 6. 

Training and validation curves of stage 2 network for tumor segmentation

Before feeding the slices to the models, we clipped the values of the input from the 0 to 2,030 range and normalize the slice with a minimum of 162 and a maximum of 2,030. Dice loss [19] was used to train all the models. We trained the lobe segmentation model for 50 epochs and performed early stopping to avoid overfitting. For lobe segmentation, in the first stage, we trained the network for 35 epochs but the network failed to converge after that, so we stopped the training because the results were good enough to continue for the second stage network. The second stage network for tumor segmentation was trained for 40 epochs and it did not show any improvements after that.

Results

Deep learning based AS model

For this part, a total of 94 cases (75 with tumors and 19 normal brain MRI scans) were included.

For each patient, MRI data of size 240 × 240 × 155 was used. We trained the model using 64 training cases, validated on 9, and tested on 21 cases. Out of 21 test cases (18 cases had tumors and 3 were normal), the AI-based model could identify tumors in 16 cases, and 2 cases were missed by the AI model. False positive cases were nil. The model didn’t predict any abnormality in normal scans or in normal slices in abnormal scans. Accuracy of detection of abnormality i.e., tumors in our trained mode was 0.90 i.e., 90%. The prediction label of a trained model is shown in Figure 7. The segmentation results for the prediction of tumor i.e., both enhancing and non-enhancing components were analyzed by Dice score and Hausdorff95 distance and mentioned in Table 3.

Display full size

Figure 7. 

Target and actual prediction labels in a case of fourth ventricle brain tumor. Figure shows prediction label of the trained model

Table 3.

Dice scores and Hausdorff95 distance of WT, ET, and NET for all datasets

Dataset	Evaluation parameters	Dice score			Hausdorff95 distance
		WT	ET	NET	WT	ET	NET
Training data (64)	Mean	0.857	0.797	0.752	1.507	2.987	3.639
	SD	0.234	0.247	0.289	1.014	4.197	3.825
	25th quartile	0.870	0.751	0.693	0.870	1.000	1.000
	75th quartile	0.952	0.957	0.985	1.560	3.000	4.030
Validation data (9)	Mean	0.712	0.507	0.437	4.328	8.475	8.502
	SD	0.136	0.285	0.215	3.181	6.862	4.596
	25th quartile	0.593	0.351	0.354	0.593	2.236	5.029
	75th quartile	0.804	0.747	0.599	7.039	13.140	11.720
Test data (21)	Mean	0.606	0.586	0.312	8.130	3.412	20.430
	SD	0.342	0.370	0.384	17.500	3.145	21.260
	25th quartile	0.573	0.198	2.95e–8	0.573	1.207	6.140
	75th quartile	0.834	0.831	0.500	5.024	4.802	20.700

Display full size

The numbers in parentheses are indicates number of examples used for training, validation and test

After training the model and segmentation, as mentioned in the materials and methodology section, all cases with tumors were subjected to heuristic rules over the deep learning-based segmentation model and features such as tumor signal intensity on T1WI, T2WI, FLAIR, enhancement pattern, cyst, necrosis, haemorrhage, and diffusion restriction were predicted for all cases. The diagnostic accuracy of feature prediction by the deep learning-based model as opposed to the radiologist detected MRI findings is shown in Table 4.

Table 4.

Diagnostic accuracy of feature prediction by deep learning based model as opposed to radiologist detected MRI findings

Predicted feature n = 70		Sensitivity	Specificity	PPV	NPV	Accuracy
Hypointensity on T1WI		86.36%	0	93.44%	0	81.43%
Isointensity on T2WI		63.16%	84.62%	94.74%	34.37%	67.14%
Hyperintensity on FLAIR		85.71%	61.9%	60%	86.67%	71.4%
Haemorrhage		54.17%	78.26%	56.62%	76.6%	70%
Cyst		79.07%	59.26%	75.56%	64%	71.4%
Necrosis		52.94%	83.02%	50%	84.62%	75.71%
Enhancement heterogeneity		81.36%	72.73%	94.12%	42.11%	80%
Enhancement quantification	Mild	50%	92.86%	82.35%	73.58%	75.71%
	Moderate	61.1%	71.15%	42.31%	84.09%	68.57%
	Severe	83.3%	84.78%	74.07%	90.70%	84.29%
Diffusion restriction		53.33%	96.77%	94.12%	68.16%	75.4%

Display full size

n: total number of abnormal cases which were picked by AS model and on which the mentioned features are predicted by AS model

Discussion

AI has recently made substantial strides in perception (the interpretation of sensory information), allowing machines to better represent and interpret complex data. Deep learning is a subset of machine learning that is based on a neural network structure inspired by the human brain. These neural networks learn discriminative features from data automatically, giving them the ability to approximate very complex nonlinear relationships. Recent methods based on deep convolutional neural networks have outperformed all traditional machine learning methods in various domains like medical image segmentation, image classification, object detection, and tracking. In the first part of the present study, clinical features and MRI based semantic features were evaluated. A total of 75 cases were included. In the second part of the study deep learning based algorithm was developed and tested to detect the abnormality in the MRI scan and segmentation into enhancing and non-enhancing components and tumor subtypes. A total of 94 cases were included in this part of the study.

The median age at diagnosis of brainstem glioma was 7 years and more common in males. A study by Hong et al. [6] revealed similar demographic features [22]. A study by Hong et al. [6] had 84% brainstem glioma cases with cranial nerve palsy involvement and 67% brainstem glioma cases with cerebellar ataxia as compared to 83.3% brainstem glioma cases with cranial nerve involvement and 94.4 % brainstem glioma cases with gait ataxia in the present study. In the present study, pilocytic astrocytoma was the second most common of these four tumors to have cranial nerve involvement (47.4%) and medulloblastoma was the second most common of these four tumors to have gait ataxia (55.6%). Univariate analysis of all tumor groups for clinical features is mentioned in the following Table 6. P-values of combined univariate analysis of all tumor groups for MRI semantic features are mentioned in Table 7.

The median age at diagnosis of ependymoma was found to be 8 years and more common in males. The mean age of ependymoma was 4.5 years according to a study by Duc et al. [23]. Ependymoma were mostly hypointense on T1WI, isointense on T2WI, and hyperintense on FLAIR. Infratentorial location was seen in 57.1% of cases. Almost all tumors (92.9%) showed heterogeneous enhancement. Hemorrhage was present in 64.3% of cases, higher than the rest of the tumor groups. Calcification was present in 35.7% of cases, relatively more common than in other groups. Diffusion restriction was present in 23.1 % of cases. Necrosis was seen in 71.4% of cases, higher than other groups. Cystic changes were seen in 71.4% of cases. Hydrocephalus was seen in 64.3 % of cases at presentation. A study by Mangalore et al. [24] with 41 cases had hydrocephalus in 34% of cases, calcification in 78% of cases and heterogeneous enhancement in all cases. However, the study by Mangalore et al. [24] mainly considered computerized tomography (CT) imaging findings as only 8 cases had baseline MRI imaging [24]. Multivariate analysis showed 16 times more odds of a tumor with necrosis having a final diagnosis of ependymoma than tumors without necrosis. It was also found that tumors with hemorrhage had 4.9 times more odds to have a final diagnosis of ependymoma than tumors without hemorrhage. Median age at diagnosis of medulloblastoma was found to be 6 years and more common in males. Median age of medulloblastoma was 9 years according to a study by Arora et al. [25]. Medulloblastoma was hypointense on T1WI and mostly isointense on T2WI and FLAIR. Midline tumor seen in 90.5 % of cases. Heterogeneous enhancement was seen in 76.2% of cases, whereas 23.8% cases showed homogeneous enhancement. Hemorrhage was present in 38.1% of cases, less than ependymoma. Calcification was seen in only 4.8% i.e., in only one case. Restricted diffusion was observed in 94.1% of cases. Cystic changes were seen in 90.5% of cases. All cases had hydrocephalus at presentation. Multivariate analysis showed 56.9 times more odds of a tumor showing restricted diffusion on DWI to have a final diagnosis of medulloblastoma than tumors not showing diffusion restriction. It was also observed that tumors with cystic changes had 25.3 times more odds to have a final diagnosis of medulloblastoma than tumors without cystic changes. Comparative analysis of a few MRI findings of medulloblastoma in the present study with two previously published studies is given in following Table 8 [26, 27].

The median age at diagnosis of pilocytic astrocytoma was found to be 9 years and more common in males. The median age of pilocytic astrocytoma was 10 years according to a study by Arora et al. [25]. Pilocytic astrocytoma was mostly hypointense on T1WI, isointense on T2WI, and hyperintense on FLAIR. Tumor was seen to be localized in forebrain in 68.4% of cases whereas 31.6% of cases were infratentorial. Heterogeneous enhancement was seen in 73.7% cases, whereas 26.3% cases showed homogeneous enhancement. Hemorrhage was present in 5.1%, less than ependymoma and medulloblastoma. Calcification was not detected in any case. Restricted diffusion observed in 23.5% of cases. Cystic changes were seen in 73.7% cases. Around 57.9% of cases had hydrocephalus at presentation. Multivariate analysis showed 13.8 times less odds of a tumor with hemorrhage to have a final diagnosis of pilocytic astrocytoma than tumors without hemorrhage.

Brainstem glioma was mostly hypointense on T1WI and isointense on T2WI and FLAIR. Heterogeneous enhancement was seen in all the cases. Haemorrhage was present in 23.8% of cases, less than ependymoma and medulloblastoma. Restricted diffusion was observed in 30% of cases. Necrosis was seen in 38.1% of cases, however, cystic changes were rarely seen (only in one case). Around 42.9% of cases had hydrocephalus at presentation. A multivariate analysis of a few significant univariate MRI features of ependymoma, medulloblastoma, and pilocytic astrocytoma is shown in Table 9.

Imaging findings were congruent with published literature by Poretti et al. [28], Plaza et al. [29], and Camacho et al. [30]. Survival analysis was not statistically significant in the present study, however findings concurrent with literature i.e., best for pilocytic astrocytoma and worst for brainstem glioma. A deep learning-based algorithm was developed and tested to detect the abnormality in the MRI scan and segmentation into enhancing and non-enhancing components. The sensitivity and specificity of this model for detecting abnormalities are 0.88 and 1 respectively (accuracy is 0.9 i.e., 90%). For the classification of paediatric brain tumors, this model showed the highest accuracy in the detection of brainstem glioma i.e., 94.2%. The accuracy of detection of pilocytic astrocytoma by this model was least i.e., 84.29%. The accuracy of detection of ependymoma and medulloblastoma was 90% and 88.57% respectively. Quon et al. [31], developed a deep learning model in a study with 617 children, which had overall classification accuracy of 92% and sensitivity of 0.96 and specificity of 1 for tumor detection [31]. However, it was a multi-institutional study. Their model was most accurate at predicting diffuse midline glioma followed by pilocytic astrocytoma and medulloblastoma [31]. Ependymoma prediction was the least accurate [31]. They considered the interpretation by four radiologists for comparison, out of which the model showed greater accuracy than two radiologists [31]. Another multi-institutional study of the deep learning-based model with 288 patients for tumor classification by Zhou et al. [32] showed an accuracy of 85% for medulloblastoma versus non-medulloblastoma, the accuracy of 80% for ependymoma versus non-ependymoma, and an accuracy of 88% for pilocytic astrocytoma versus non-pilocytic astrocytoma [32] as shown in Table 10. It had significantly higher accuracy than the average qualitative expert MR imaging review [32].

Our study shows that the prediction of tumor types by experienced radiologists was more accurate than deep learning model prediction. This primarily is related to the fact that a smaller number of cases were included in this study for training and validation purposes. The accuracy of the model can be increased further in a greater number of cases. Nevertheless, the model may form the basis of molecular genetics prediction by deep learning of the most common paediatric brain tumors.

In conclusion AI has the potential in localizing paediatric brain tumors and feature detection as well as diagnosis. The proposed deep learning-based model had very high accuracy and specificity for predicting the individual MR characteristics and close to 80% accuracy in predicting tumor type. This model can serve as a potential tool to aid to make timely and accurate diagnosis for radiologist not specialized/trained in neuroradiology and neuro-oncology. Nevertheless, the model may also form the basis of a multi-omics (clinical radiological and pathological) model for predicting molecular genetics by deep learning/machine learning.