Abstract
Glioma is common type of brain tumour in adults originating from glia cell. Despite advances in medical image analysis and gliomas research, accuarte diagnosis remains a challenge. Gliomas can be in general classifed into High Grade (HG) and Low Grade (LG). The exact classification of glioma helps in evaluating the disease progression and selection of the treatment strategy. Whilst medical image classification using a Convolutional Neural Networks (CNNs) has achieved remarkable success, but it is still difficult task for CNNs to accurately classify 3D medical images. One of the major limitation is the fact that CNNs are difficult to optimize in 3D volumetric classification. In current work, we addressed this challenge by introducing a cascade of CNN with Long Short Term Memory (LSTM) Network for classification of 3D brain tumor MR images into HG and LG glioma. Features from pre-trained VGG-16 were extracted and fed into LSTM network for learning high-level feature representations to classify the 3D brain tumour volumes into HG and LG glioma. The results showed that the features extracted from VGG-16 gave better classification accuracy as compared to the features extracted from AlexNet and ResNet.
I. INTRODUCTION
Abnormal growth of cells in the brain results in a mass, which pushes on the normal structures in the brain. This abnormal mass grows quickly and creates a mass of abnormal cells called a tumour. Brain tumours vary in shape, size and severity level. They are heterogeneous in nature i.e. they can occur anywhere in the Central Nervous System (CNS) and have different Image Intensities. The most frequent brain tumours in adults are gliomas. The abnormal growth of the glial cell in the brain that surrounds the neurons results in the formation of gliomas[1].
V. CONCLUSION & FUTURE WORK
In this work, we have proposed a cascaded CNNLSTM model for volumetric classification of a brain tumour into High Grade and Low Grade glioma. Experimental results on the state of the art CNN architectures have shown that the VGG-16 model performs the best by extracting the high-level feature representations and thus enables the LSTM to effectively discriminate between HG and LG glioma. Another advantage of this method is that it is able to perform robustly on high dimensional 3D data. In future, we will extend this work for segmentation task by designing a model that can accurately predict the pixel labels for a brain tumour with a limited amount of training data.
