The design of 3D analysis-based framework for glioma characterization in MR images

Computed Tomography (CT) and Magnetic Resonance (MR) imaging techniques are frequently used-scanning methods for the detection of tumors. MR imaging is more frequently used than other imaging modalities to highlight brain abnormalities since it can detect even a small structure in the brain tissue. In the literature, the discrimination of High-Grade Glioma (HGG) and Low-Grade Glioma (LGG) is performed using semi-automated approaches. In other words, the classification process is provided by examining Two Dimensional (2D) image analyses-based models. At this point, a fully-automated Computer-Aided Diagnosis (CAD) system can only be realized by handling the tumor in Three Dimensional (3D) and by designing an applicable – task-based classification framework. In the thesis, the classification part of a fully automated CAD is considered, and a promising model is suggested to grade the gliomas on a 3D basis. In the proposed model, all phase information (T1, T2, T1c, FLAIR) is evaluated in 3D MR images, and a 3D to 2D Feature Transform Strategy (3t2FTS) is offered to form the input data. In 3t2FTS, space transform is provided with the usage of First-order Statistics (FOS). Herein, the main purpose is to present the input information to be fed into the transfer learning architectures by using an efficient approach. Concerning this, 2D-ID images which are identifiers for tumors, are obtained by converting the 3D voxel information to 2D images. In our work, these images are formed for every tumor, and the data is tested on eight transfer learning algorithms (DenseNet201, InceptionResNetV2, InceptionV3, ResNet50, ResNet101, SqueezeNet, VGG19, Xception). To detect the most remarkable deep learning architecture and to examine the performance of the proposed model, the BraTS 2017/2018 dataset is evaluated on discrimination of 210 HGG and 75 LGG samples. Hyperparameters of architectures are comprehensively analyzed to reveal the highest performance to be reached. 2-fold cross-validation is chosen as the test method to evaluate the system performance. Consequently, it's observed that the framework operating 3t2FTS and ResNet50 can perform the HGG – LGG discrimination by achieving 80% classification accuracy. Regarding this, it's revealed that the achieved success score and the proposed 3t2FTS-based classification are promising in terms of progress.