Introduction

Accurate delineation of brain tumor is important for diagnosis and treatment planning by the radiologist. Magnetic Resonance Imaging(MRI) is a widely used medical imaging technique for diagnosis of brain tumor. On contrast-enhanced T1-weighted(W) MR images, tumor area where Blood-Brain Barrier breaks show enhancement and it is an indication of an aggressive type of tumor. It is necessary to segment the active or aggressive tumor region for better diagnosis and surgical planning. Various manual, automatic and semi-automatic segmentation techniques are reported in literature for brain tumor segmentation^1,2,3,4. Manual delineation of tumor region is a time-consuming process and radiologist dependent thus leads to intra and inter variability in segmentation results⁵. So, there is a need to automate the process of brain tumor segmentation. The current study aimed to develop Support Vector Machine(SVM) classifier based semi-automatic segmentation of contrast-enhanced(CE) brain tumor region. We hypothesize that for SVM based CE tumor region segmentation, the relative percentage image obtained from T1-W pre and post-contrast MR data provides better accuracy as compared to conventional post-contrast T1-W images.

Method

In this IRB-approved retrospective study, MRI data of 20 brain tumor patients(4 Grade II, 8 Grade III and 8 Grade IV patients) have been acquired at a local hospital on 3T whole-body MRI (Ingenia, Philips Healthcare, The Netherlands). The MRI protocol for this study includes conventional T1-W, post-contrast T1-W and Dynamic Contrast-Enhanced(DCE) T1-W MR images. The MRI protocol parameters are shown in Table 1. Pre-processing(Noise Removal, Skull Stripping and Registration) of post-contrast T1-W and DCE MRI data was done. Conventional post-contrast T1-W MR images were normalized between 0-255. Relative percentage image(RPI) was calculated by taking the difference of average of last 4 time-point(Spre) from the average of first 4 time-point images(Spost) of DCE MRI data and further dividing it by Spre. The obtained relative image was normalized between 0-255 and named as RPI. CE tumor region mask was manually segmented by MTech student and validated by experienced Radiologist(>25 years of MR imaging experience). SVM classifier was trained on the manually segmented mask for CE region on both RPI and post-contrast T1-W images separately for 15 patient datasets(3 Grade II, 6 Grade III and 6 Grade IV patients). SVM classifier was implemented using MATLAB 2018b and optimized with respect to its hyperparameters(c for linear kernel; c and gamma for Radial Basis Function (RBF) kernel) based upon 10-fold cross-validation. Trained SVM classifier was tested on additional 5 brain tumor patients(1 Grade II, 2 Grade III and 2 Grade IV) for segmentation of accurate CE tumor region. Rather than using an entire image for training/testing, a local region-of-interest(ROI) around tumor region in slice containing maximum enhancing tumor lesion was selected and further processing was carried out only within this ROI. This manual ROI was drawn on a single central tumor slice of the patient and multiplied with all the slices to get the total volume of the CE tumor. The RPI intensity values and post-contrast T1-W image intensity values of the ROI region are used as feature set for training SVM classifier separately to segment CE tumor region.

Results and Discussion

Figure 1 shows the curve of the training and validation error at different c values obtained while training a linear SVM classifier. The graph shows that the least validation error(1.078%) is obtained at c = 0.1 for RPI. In the case of post-contrast T1-W image-based SVM training, the least validation error of 23% is obtained at c = 0.001. Test error of SVM classifier at best c value was 3.26% in case of RPI and 25% in case of post-contrast T1-W image. Figure 2 shows the semi-automatic SVM based CE tumor region segmentation results using RPI for 3 tumor patients, one from each grade. The SVM based segmentation using post-contrast T1-W image have high error as compared to RPI based SVM segmentation due to mixture of several tissue signals in a single voxel in post-contrast T1-W image causing partial volume effect. Also, Enhancing tumor area on post-contrast T1-W image may represent false-enhancing non-tumor component that may be an infarct. The segmentation results using post-contrast T1-W image include these enhancing non-tumor components also if present but the proposed approach using RPI removes these components from enhancing tumor region as shown in Figure3. Figure4 shows the signal intensity time curve for the enhancing tumor region and the infarct region. The current approach needs slight modification for the removal of blood vessels which might be present in the tumor region. The current study validates that the proposed SVM based semi-automatic segmentation approach can be used to segment CE tumor region using RPI. However, the study needs to be validated on more number of datasets and further modification in the proposed approach needs to be done.

Conclusion

SVM based upon RPI intensities provided accurate segmentation of CE tumor compared to conventional post-contrast T1-W based SVM classifier. The proposed method accurately removed infarct region from segmented CE tumor region, which is a challenge in post-contrast T1-W images. The proposed approach require selection of a local region-of-interest around tumor region in a single slice containing largest tumor area. This ROI approach reduced processing time and improved accuracy.

Acknowledgements

This work was supported by Indian Institute of Technology, Delhi and Fortis Memorial Research Institute, Gurugram. The authors thank Dr. Anirban Sengupta and Dr Sumeet Agarwal for discussion on machine learning.