A hierarchical Convolutional Neural Network for Segmentation of Stroke Lesion in 3D Brain MRI

Introduction: Brain tumors such as glioma are among the most aggressive lesions, which result in a very short life expectancy in patients. Image segmentation is highly essential in medical image analysis with applications, particularly in clinical practices to treat brain tumors. Accurate segmentation of magnetic resonance data is crucial for diagnostic purposes, planning surgical treatments, and also follow-up evaluation. Manual segmentation of a large volume of MRI data is a time-consuming endeavor, and this necessitates employing automatic segmentation techniques that are both accurate and reliable. However, the vast spatial and structural diversity of brain tissue poses serious challenges for this procedure. The current study proposed an automatic segmentation method based on convolutional neural networks (CNN), where weights of a pre-trained network were used as initial weights of neurons to prevent possible overfitting in the training phase. Methods: As tumors were diverse in their shape, size, location, and overlapping with other tissue, it was decided to exploit a flexible and extremely efficient architecture tailored to glioblastoma. To remove some of the overlapping difficulties, morphological operators as a pre-processing step were utilized to strip the skull. Results: The proposed CNN had a hierarchical architecture to exploit local and global contextual features to handle both high- and low-grade glioblastoma. To address biasing stem from the imbalance of tumor labels, dropout was employed and a stochastic pooling layer was proposed. Conclusions: Experimental results reported on a dataset of 400 brain MR images suggested that the proposed method outperformed the currently published state-of-the-art approach in terms of various image quality assessment metrics and achieved magnitude fold speed-up.