Automatic brain tumor segmentation from multimodal MRI plays a significant role in assisting the diagnosis, treatment, and surgery of glioblastoma and lower glade glioma. In this article, we propose applying several deep learning techniques implemented in AWS SageMaker Framework. The different CNN architectures are adapted and fine-tuned for our purpose of brain tumor segmentation.The experiments are evaluated and analyzed in order to obtain the best parameters as possible for the models created. The selected architectures are trained on the publicly available BraTS 2017–2020 dataset. The segmentation distinguishes the background, healthy tissue, whole tumor, edema, enhanced tumor, and necrosis. Further, a random search for parameter optimization is presented to additionally improve the architectures obtained. Lastly, we also compute the detection results of the ensemble model created from the weighted average of the six models described. The goal of the ensemble is to improve the segmentation at the tumor tissue boundaries. Our results are compared to the BraTS 2020 competition and leaderboard and are among the first 25% considering the ranking of Dice scores.
1 August 2018 in Acta Universitatis Sapientiae, Informatica
Several hundreds of thousand humans are diagnosed with brain cancer every year, and the majority dies within the next two years. The chances of survival could be easiest improved by early diagnosis. This is why there is a strong need for reliable algorithms that can detect the presence of gliomas in their early stage. While an automatic tumor detection algorithm can support a mass screening system, the precise segmentation of the tumor can assist medical staff at therapy planning and patient monitoring. This paper presents a random forest based procedure trained to segment gliomas in multispectral volumetric MRI records. Beside the four observed features, the proposed solution uses 100 further features extracted via morphological operations and Gabor wavelet filtering. A neighborhood-based post-processing was designed to regularize and improve the output of the classifier. The proposed algorithm was trained and tested separately with the 54 low-grade and 220 high-grade tumor volumes of the MICCAI BRATS 2016 training database. For both data sets, the achieved accuracy is characterized by an overall mean Dice score > 83%, sensitivity > 85%, and specificity > 98%. The proposed method is likely to detect all gliomas larger than 10 mL.
In this paper we propose to create an end-to-end brain tumor segmentation system that applies three variants of the well-known U-Net convolutional neural networks. In our results we obtain and analyse the detection performances of U-Net, VGG16-UNet and ResNet-UNet on the BraTS2020 training dataset. Further, we inspect the behavior of the ensemble model obtained as the weighted response of the three CNN models. We introduce essential preprocessing and post-processing steps so as to improve the detection performances. The original images were corrected and the different intensity ranges were transformed into the 8-bit grayscale domain to uniformize the tissue intensities, while preserving the original histogram shapes. For post-processing we apply region connectedness onto the whole tumor and conversion of background pixels into necrosis inside the whole tumor. As a result, we present the Dice scores of our system obtained for WT (whole tumor), TC (tumor core) and ET (enhanced tumor) on the BraTS2020 training dataset.