4.4.1 SP_ERT Single Modality Method
FLAIR images are commonly acquired in clinical practice as part of standard diagnostic clinical MRI of brain tumours. The experimental results in this chapter and Section 4.3.4, which are shown in Table 4-6, Table 4-7, Figure 4-25 and Figure 4-26, 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 2013 training dataset (FLAIR) with the similar model parameters and features tuned for the clinical dataset. The experimental results in Section 4.3.5, which are shown in Table 4-8 and Table 4-9, suggests the robustness of the SP_ERT single modality method.
Selecting an appropriate superpixel size is essential for increasing the overall segmentation accuracy within an optimum calculation speed. Selecting large superpixel size requires fewer number of total superpixels, hence it can ensure fast computation and meanwhile may provide sufficient information for feature extraction such as stable texture features. On the other hand, a large superpixel size may contain more than one class of pixel which may cause inaccurate feature calculation (such as small areas of calcification or haemorrhage), and it is also not
108
appropriate for small sized lesions. Whereas, a small superpixel size has higher probability to purely contain one class of pixel, hence it is preferred for segmentation of a small lesion. However, they may not be enough pixels for calculating stable features. Also, the computation time for generating the small size partitions is very high due to the large number of total superpixels. The aim of optimisation is to find a superpixel size which provides a good trade- off between computation time and segmentation accuracy. In Section 4.3.3, the size of superpixel is obtained through exhaustive parametric searching during the training stage on the selected data.
The compactness factor is another important parameter for superpixel segmentation which should be tuned. As explained in Section 4.3.3, higher value of compactness factor provides more rigid partitions which are more stable and usually less noisy. In the case of superpixel segmentation noise is considered as holes or sparse separated pixels. However, the rigid superpixels may not follow the tissue boundaries very well, especially in the cases where there are no sharp or clear boundaries. On the other hand, lower values for the compactness factor provides more flexible and accurate boundaries, but produces more isolated and disconnected pixels. They also may generate very narrow superpixels which are not appropriate for texture analysis. In the Section 4.3.3, the compactness factor is determined using visual inspection of matching the superpixels with the boundaries of the ground truth.
The application of SP_ERT method on the BRATS data is compared to the methods published in (Menze et al., 2015) which used the same data in MICCAI challenge. However, some of the corresponding methods are assessed on the training dataset, whilst others are on the separate testing dataset. Since the current study is based on binary classification (i.e. tumour including oedema and active tumour core versus normal brain tissue) using a single FLAIR protocol, it is difficult to have a direct comparison with the current published methods on BRATS data. However, the experimental results of the SP_ERT method are in the same range of other methods and are close to the best segmentation of the complete tumour which demonstrates the strength of the method. The image patches from the superpixels are following the edges in the images, therefore including more homogeneous pixels. This increases the robustness of the final segmentation after classification of superpixels. However, in the case of small tissues that encompass few superpixels or those smaller than an average superpixel, the algorithm might fail. Misclassification of these superpixels will result in assigning the small tumorous region as healthy brain tissues.
This study also emphasises the importance of MRI histogram normalisation in the preprocessing stage. This is of importance especially when the method is applied to the data that are from multi-centres and different scanners such as BRATS dataset. When the histogram
109
normalisation was applied on the clinical data, there were only slight differences for the Dice scores, for the individual patient data, whereas the mean Dice score for all the 19 data is the same as before (i.e. 0.91). This is mainly because the clinical data are quite consistent.Also, the feature normalisation step was found out to be quite important, especially when it was applied to the BRATS data, the segmentation results improved significantly using both SP_SVM and SP_ERT. This may be the reason why even without histogram normalisation, the mean Dice score of 0.82 for SP_ERT was still obtained (partial normalisation) as shown in Table 4-2.
4.4.2 Applying the SP_ ERT on BRATS dataset
The BRATS training dataset was used to evaluate the robustness of the method. As discussed in the Section 4.3.5, most of the parameters are the same as those optimised for the clinical data. Figure 4-33 shows the overall average and standard deviation of Dice score overlap measures for all 19 clinical patient data and 30 BRATS 2013 dataset using both SP_ERT and SP_SVM methods. The results show that using the state-of-the art ERT for classification of superpixels provides more accurate and robust segmentation compared to that of an SVM classifier. For the clinical dataset, the Dice score overlap measure for SP_ERT segmentation is 0.91 ± 0.04, while for SP_SVM method it is 0.87 ± 0.05. For BRATS 2013 dataset, the Dice score overlap measure for SP_ERT segmentation is 0.88 ± 0.05, while for SP_SVM method it is 0.83 ± 0.06. The mean Dice scores obtained from BRATS training dataset is closer to that from the clinical dataset, this suggests robustness of the method.
110
Figure 4-33 Comparison of the average and standard deviation of Dice score overlap measures for SP_SVM vs. SP_ERT for all 19 data scans in the clinical dataset and 30 clinical scans in BRATS 2013 dataset.
Table 4-10 presents the comparison of applying the SP_ERT single modality method on BRATS 2013 dataset with the best scores in the MICCAI challenges (Menze et al., 2015). The method proposed by Tustison et al. (Tustison et al., 2013) was the winner of on-site BRATS 2013 challenge and performed on the challenge data. The best on-site score could provide a comparable reference using BRATS dataset despite the difference between datasets. The SP_ERT single modality method was also compared to the method proposed by Reza et al.
(Reza and Iftekharuddin, 2013) which has the best result for the training set of the BRATS multi-protocol dataset. This is the same dataset which was used in the experiments of the SP_ERT method, however only the FLAIR protocol was used. This work has achieved the average Dice overlap of 0.88 which is closer to that of 0.92 by Reza’s method. As explained in Section 4.3.5 to emphasise the robustness of the SP_ERT method, the similar optimum parameters and the same five features selected for the clinical dataset are directly applied to the BRATS dataset. The algorithm is trained particularly on 1.5T clinical data from a single centre, whereas the BRATS data contains multicentre data from 1.5T and 3T MRI scanners. This may be the reason for the slightly difference of the results between the two datasets.
111
Table 4-10 Comparison with other related methods using BRATS dataset (MICCAI 2013). Note: the proposed SP_ERT method and Reza et al. (Reza and Iftekharuddin, 2013) are performed on BRATS clinical training data and the other work (Tustison et al. (Tustison et al., 2013)) is performed on BRATS challenge data.