Brain Tumor Segmentation Based on Zernike Moments, Enhanced Ant Lion Optimization, and Convolutional Neural Network in MRI Images

Gliomas that form in glial cells in the spinal cord and brain are the most aggressive and common kinds of brain tumors (intra-axial brain tumors) due to their rapid progression and infiltrative nature. The procedure of recognizing tumor margins from healthy tissues is still an arduous and time-consuming task in the clinical routine. In this study, a robust and efficient machine learning-based pipeline is suggested for brain tumor segmentation. Moreover, we employ four MRI modalities for increasing the final accuracy of the segmentation results, namely, Flair, T1, T2, and T1ce. Firstly, eight feature maps are extracted from each modality using the Zernike moments approach. The Zernike moments can create a feature map using two parameters, namely, n and m. So, by changing these values, we are able to generate different sets of edge feature maps. Then, eight edge feature maps for each modality are selected to produce a final feature map. Next, four original images are encoded into new four images to represent more unique and key information using the Local Directional Number Pattern (LDNP). As different encoded image leads to obtaining different final results and accuracies, the Enhanced Ant Lion Optimization (EALO) was employed to find the best possible set of feature maps for creating the best possible encoded image. Finally, a CNN model is utilized to explore significant details from the brain tissue more efficiently which accepts four input patches. Overall, the suggested framework outperforms the baseline methods regarding Dice score and Recall.