Fig 1.
Examples of the BraTS 2020 dataset on brain tumors images.
Yellow: enhancing tumor (ET), Blue: non-enhancing tumor/necrotic tumor (NET/NCR), Green: peritumoral edema (ED) [4].
Fig 2.
Overall block diagram of the proposed strategy used for enhancing brain tumor segmentation accuracy.
Fig 3.
Proposed minimax normalization.
Fig 4.
Proposed combined techniques for data augmentation.
Fig 5.
Overview of proposed neural network architecture.
Fig 6.
A visualization of how max pooling works and its role in spatial down sampling within the 3D convolutional neural network architecture, each colored moving window captures the maximum value inside the 2x2x2 cube and outputs it on the right-hand side.
Fig 7.
Illustration of the application of the proposed related-pooling function.
Fig 8.
Segmentation Results of Eight BraTS’2020 Validation Cases: Image Modalities, Ground Truth Labels, and Ensemble Model Predictions.
Table 1.
Model assembly (snapshot ensemble) and validation performance on BraTS 2020.
Pipelines A and B are trained independently with different normalization settings; the ensemble averages probability maps from the selected epochs per pipeline.
Table 2.
Additional information to visualize the performance of each model over the last 50 epochs.
Fig 9.
Plot of the first validation loss with Dice Score Metrics of the 3D-UNET for 300 epochs with Pipeline A and B.
Table 3.
Comparison of different approaches by selecting the best model epochs on the BraTS 2020 validation set.
Table 4.
Comparison of different approaches’ performances on the validation set (means; best in bold).
Fig 10.
Performance comparison of proposed method with state-of-the-art techniques on BraTs 2020 dataset in terms of dice coefficient and Hausdorff Distance.
Table 5.
Model performance on an internal test set extracted from the BraTS 2020 training dataset (means; best in bold).
Fig 11.
Visual segmentation results of proposed ensemble model method.
From left to right, show the axial slice of MRI images in two modalities, ground-truth and predicted results.