Automated brain tumour detection and segmentation using superpixel-based extremely randomized trees in FLAIR MRI

Despite improvements in the diagnosis and oncological treatment of primary brain tumours, they remain associated with significant morbidity and a poor overall prognosis. The majority of primary brain tumours originate from glial cells (termed glioma) and are classified by their histopathological appearances using the World Health Organization (WHO) system into low-grade glioma (LGG) (grades I and II) and high-grade glioma (grade III anaplastic glioma and grade IV glioblastoma). The typical natural history of low-grade glioma is a latent period of growth and infiltration of white matter with subtle neuro-cognitive deficit and seizures in some cases followed by regional change or transformation to a more malignant variant. High-grade glioma may present as a de novo (primary) glioblastoma or as a transformation of a lower-grade tumour (e.g. secondary glioblastoma).

Gliomas typically originate within white matter and exhibit irregular growth patterns along white matter fibres, infiltrating surrounding brain. As a result, they exhibit irregular boundaries that may be visually indistinct on conventional magnetic resonance images. Delineation of the tumour boundary and assessment of tumour size are needed for patient management in terms of treatment planning and monitoring treatment response, and current guidelines incorporate the use of both contrast-enhanced T1-weighted (CE T1w) images and T2-weighted (T2w) / FLAIR images [1, 2]. Many low- grade gliomas do not show contrast enhancement; hence, T2w/FLAIR images are used to define the tumour extent and volume. A longitudinal study has shown that LGG volume and growth rate can be used to assess whether patients are at risk with tumours likely to undergo an early malignant transformation [3]. In clinical studies, current Response Assessment in Neurooncology (RANO) criterion simply uses a bidirectional measurement to determine tumour volume for assessing treatment response [4]. Although a full 3D volume measurement may provide a more accurate volume assessment, there is a need for accurate and fully automated methods since manual segmentation (region of interest drawing) around tumour margins on a slice-by-slice basis is time-consuming and can take 12 min or more per tumour, with semiautomatic methods taking 3–5 min [5, 6]. T2w/FLAIR images can also be useful to help define the target volumes for radiotherapy planning of high-grade gliomas [2, 5]; hence, an automated segmentation that is not subject to operator subjectivity may be beneficial [5]. In this study, we have concentrated on developing and validating an automated method for a single MRI modality, FLAIR, that could be readily translated for clinical use. Future automated methods are likely to incorporate information from multimodal clinical MRI as in the Multimodal Brain Tumor Image Segmentation Benchmark (BRATS) database studies [7‐9] and also include perfusion and diffusion imaging to detect tumour tissue subtypes (e.g. necrosis, active tumour, infiltrative tumour, oedema) [10].

However, automated detection and segmentation of brain tumour is a very challenging task due to its high variation in size, shape and appearance (e.g. image uniformity and texture) [11]. Also, typical clinical image acquisition protocols usually lead to higher intraslice resolution than interslice resolution to achieve the balance of good apparent image resolution with adequate signal to noise and restricted scanning time that causes asymmetry in partial-volume effects. High-grade gliomas usually have irregular boundaries which, in some cases, are unclear or discontinuous [12]. Current work on brain tumour segmentation can be categorized into atlas-based [13‐15], unsupervised [16‐19], hybrid [20‐22] and supervised- based approaches [23‐26].

In Ref. [20], a hybrid method was proposed for brain tissue detection in MRI images which included seeded region growing segmentation and neural network classification. However, the method was semiautomatic and different parts of tumour need to be pre-selected to initiate the segmentation process. Another method is proposed for detection of multiple sclerosis (MS) lesions in brain MR images which consisted of rule-based, level-set and support vector machines [21]. Rajendran and Dhanasekaran [22] proposed a hybrid method for segmenting the tumour by combining region-based fuzzy clustering and deformable model. However, the method was only applied on a few FLAIR images with fixed parameters.

Supervised learning-based algorithms use training data labelled by experts for segmentation of tumours. Geremia et al. [23] used discriminative random decision forests to classify the voxels of 3D MRI image for segmentation of MS. Wu et al. [24] used superpixel features in a conditional random fields (CRF) framework to detect brain tumours. However, the method was not satisfactory for low-grade tumours segmentation. A method was proposed in [25] which used extremely randomized forest classification considering both appearance and context-based features. Another method was proposed in [26] which used ERT for classification of voxels based on their statistical and textural features, which were extracted using different MRI protocols. In order to reduce the computation time, it was suggested that features were only extracted from a random set of voxels, but this resulted in losing some part of data. In addition, a fixed size neighbourhood for each voxel was used to calculate features.

A number of advanced algorithms [23, 27‐31] were recently presented in [7] using the BRATS [8, 9] organized in conjunction with the international conference on Medical Image Computing and Computer-Assisted Interventions (MICCAI) 2012 and 2013 conference. The methods were based on segmentation of different tumour tissues, i.e. tumour core, oedema, necrosis, using multiprotocol containing FLAIR, T1-weighted (T1w), T1w with contrast and T2-weighted protocols [32].

Despite much effort being devoted to the segmentation problem, brain tumour segmentation remains an ongoing research topic. Very few completely automatic segmentation algorithms have been adopted in the clinic. Recently, only one automated tool has been clinically evaluated [33].

In this study, we investigate a fully automated superpixel-based method for detection and segmentation of the abnormal tissue associated with brain tumours as defined by the \(\hbox {T}_{2}\) hyperintensity from Fluid-Attenuated Inversion Recovery (FLAIR) MRI. FLAIR images are routinely acquired as part of standard diagnostic clinical MRI of brain tumours. Delineation of the FLAIR hyperintensity is relevant for assessing low-grade glioma growth [34], defining an abnormality region from which imaging features for tumour classification can be extracted [35], aiding with radiation dose planning [36] and assessing treatment response [37]. Different from the methods in [25] and [26], in which image features were calculated based on each individual voxel and a fixed size neighbour-hood was considered for the feature extraction, in this paper, superpixel partition is firstly calculated which provides accurate boundaries between different tissues, and then image features are extracted from each superpixel. This will not only improve the accuracy of feature calculation, but also increase the speed of computation. We demonstrate the automated method that provides a close match to expert delineation across all grades of glioma and so could provide a faster and more reproducible method of brain tumour delineation to aid patient management. To assess the robustness of the proposed method, the method is also evaluated on the FLAIR protocol of BRATS 2012 annotated training dataset [8, 9].

The rest of this paper is organized as follows. The “Method” section describes the proposed method, including superpixel partition, feature extraction, classification and final segmentation. The “Experimental results” section presents the data description and experimental results for the two datasets, followed by “Discussion” and “Conclusion” sections.

Our method consists of four main steps, which are depicted in Fig. 1. After preprocessing, in the superpixel segmentation step, FLAIR image is partitioned into irregular patches with approximately similar size and intensity values. For each superpixel patch, a number of features including statistical, texton and shape features are calculated. This is then followed by feature selection to find the most significant features, based on which each superpixel is classified into tumour and non-tumour using an ERT classifier.

Two experiments were carried out in this section. In the first experiment, our own clinical dataset is used for training and validation of the algorithm. In the second experiment, the method is further validated on the publicly available MICCAI BRATS 2012 dataset [7‐9] to assess the robustness of the method. The following subsections, including data description, parameters selection and comparative experimental results are focused on our own data cohort, while the next subsection presents the evaluation results on MICCAI BRATS 2012 clinical training dataset.

FLAIR images are routinely acquired in clinical practice as part of standard diagnostic clinical MRI of brain tumours. Our experimental results shown in Tables 4, 5 and Fig. 10 demonstrate high performance of automated detection and segmentation of the brain tumour oedema and core regions in FLAIR MRI. The method was also further validated on BRATS 2012 training dataset (FLAIR) with the similar model parameters and features tuned for our own clinical dataset; good results shown in Tables 6 and 7 suggest the robustness of our method.

Selecting an appropriate superpixel size is critical for increasing the overall segmentation accuracy within an optimum calculation speed. Large superpixel size can ensure fast computation and may provide sufficient information for feature extraction such as stable texture features. However, large size of superpixel may contain more than one class of pixels which leads to inaccurate feature calculation (such as small areas of calcification or haemorrhage), and it is also not suitable for small-sized lesions. While small size of superpixel has higher probability of purely containing one class of pixels, it is preferred for small lesion segmentation. However, they may not have enough pixels for calculating stable features, and the computation time for generating the small size partitions is very high. In this study, the size of superpixel is obtained through exhausted parametric searching during the training stage. An optimization algorithm such as genetic algorithm can be explored to effectively find an optimum superpixel size which provides a good trade-off between computation time and segmentation accuracy.

Another important parameter in superpixel segmentation stage is the compactness factor. Higher value of this parameter leads to more rigid partitions which are more stable and usually less noisy, i.e. holes or sparse separated pixels. However, the segmentation may not follow the tissue boundaries very well, especially in the cases where there are no sharp or clear boundaries. While, lower compactness values result in more flexible and accurate boundaries, but the segmentation may produce more isolated and disconnected pixels. They also may generate very narrow superpixels which are not appropriate for texture analysis. In our current study, the compactness factor is determined through visual inspection. Optimization methods need to be investigated to obtain an optimum compactness factor which provides a good trade-off between noise and flexibility.

For the comparison of our method on BRATS data, we refer to the work published in [7] which used these data in MICCAI challenge. However, some of their methods are assessed on the training dataset, while others are on the separate testing dataset. Due to the fact that our current study is based on binary classification (i.e. tumour including oedema and active tumour core versus normal brain tissue) using single FLAIR protocol, it is difficult to have a direct comparison with the current published methods on BRATS data. However, our results which are in the same range of other methods and are close to the best segmentation of whole tumour demonstrate the promise of the method.

Although currently we have only evaluated our segmentation algorithm for FLAIR images, it should be straightforward to apply the same superpixel methodology to contrast-enhanced T1w images and determine the signal intensity and higher-order features that best segment the contrast-enhancing region of high-grade gliomas. In fact, we are currently working on this direction.

In this study, we note the importance of the preprocessing step, namely MRI histogram normalization. This is of particular importance when the method is applied to BRATS dataset, whose data are from multicentres and different scanners.

In the current study, we also note in Fig. 13a2 that small hypointense spots in the FLAIR (and corresponding T1w) may be calcifications, and the hypointense FLAIR region, which is excluded by the SVM method (Fig. 13c2) but included in the ERT analysis (Fig. 13d2), is haemorrhagic since there is hyperintensity in the T1w MRI. This is a limitation of the current single modality analysis if these regions need to be separately specified. Future studies extending our method to multimodal data are planned. This will include the segmentation of different tissue subtypes (e.g. necrosis, active tumour, infiltrative tumour, oedema) by incorporating information from multimodal clinical MRI, including perfusion and diffusion imaging.

For our own clinical dataset, the ground truths were provided based on one expert’s manual annotation. There may have some errors in the manual annotations, which may include intratumoural bleeding or calcification in the tumour (e.g. in Fig. 13b2). When those annotations are used to train the model, it may lead to some errors in the final segmentation. Also, our current clinical dataset mainly contains general cases, such as different tumour grades from a wide range of patient ages (patient ages at the time of scanning ranged from 22 to 73). In the future, we will look into more complicated cases, such as calcification, intratumoural bleeding or elderly patients with white matter disease, which are clinically very important to distinguish against.

This paper proposed a fully automated method for the detection and segmentation of brain tumour from FLAIR MRI images, by calculating Gabor texton feature, fractal analysis, curvature and statistical intensity features from superpixels. ERT is then used to classify each superpixel into tumour or healthy brain tissue. The formation of superpixel by grouping voxels with similar properties and extracting features from superpixels can not only improve the accuracy of feature extraction, especially for the superpixels near the boundaries between different tissues, but also significantly reduce the computation time, compared to voxel-based feature calculation and classification. The experimental results demonstrate the high detection and segmentation performance of the proposed method using ERT classifier, with average sensitivity of 89.48 %, BER of 6 % and Dice overlap ratio of 0.91. To assess the robustness of the method, the method was further evaluated on BRATS 2012 dataset, which results in similar good performances of 88.09 %, 6 % and 0.88, respectively. This provides a close match to expert delineation across all grades of glioma, leading to a faster and more reproducible method of brain tumour delineation to aid patient management.