Source-linked AI summary
Segmentation of Glioma Tumors in Brain Using Deep Convolutional Neural Network
Saddam Hussain, Syed Muhammad Anwar, Muhammad Majid
TL;DR
Glioma segmentation is challenging because tumor regions and tissues have irregular shapes, while some context-aware approaches are computationally expensive. The paper proposes an ILinear nexus architecture and reports improved performance over state-of-the-art techniques, with strong results for core and enhancing regions.
Problem
Glioma segmentation is challenging because tumor regions and tissues have irregular shapes, requiring large portions of the image to be considered for accurate segmentation.
Method
The paper stacks two networks to form an ILinear nexus architecture and uses improved N4ITK bias field correction during preprocessing.
Results
The proposed architecture improves performance compared with state-of-the-art techniques and produces outstanding results for core and enhancing regions.
Takeaways & Limitations
The architecture achieves its strongest reported results for core and enhancing tumor regions, while whole-tumor performance is weaker.
Takeaways & Limitations
The proposed approach has a limitation involving the labeling of effective edema features as the major tumor class, and context-aware approaches can be computationally expensive.
Abstract
from arXiv · showhide
Detection of brain tumor using a segmentation based approach is critical in cases, where survival of a subject depends on an accurate and timely clinical diagnosis. Gliomas are the most commonly found tumors having irregular shape and ambiguous boundaries, making them one of the hardest tumors to detect. The automation of brain tumor segmentation remains a challenging problem mainly due to significant variations in its structure. An automated brain tumor segmentation algorithm using deep convolutional neural network (DCNN) is presented in this paper. A patch based approach along with an inception module is used for training the deep network by extracting two co-centric patches of different sizes from the input images. Recent developments in deep neural networks such as drop-out, batch normalization, non-linear activation and inception module are used to build a new ILinear nexus architecture. The module overcomes the over-fitting problem arising due to scarcity of data using drop-out regularizer. Images are normalized and bias field corrected in the pre-processing step and then extracted patches are passed through a DCNN, which assigns an output label to the central pixel of each patch. Morphological operators are used for post-processing to remove small false positives around the edges. A two-phase weighted training method is introduced and evaluated using BRATS 2013 and BRATS 2015 datasets, where it improves the performance parameters of state-of-the-art techniques under similar settings.
1. Introduction
Glioma segmentation is difficult because tumors have irregular structures, fuzzy boundaries, heterogeneous MRI intensities, and severe pixel-level class imbalance. The paper proposes an automated CNN-based method that combines contextual and local features while targeting computational efficiency.
- MRI-based tumor detection is complicated by scanner configurations and intensities that vary across voxels and can mask tumor regions.
- Multi-modal MRI combines structural and functional information that can improve localization of irregularly shaped tumors compared with a single modality.
- Gliomas are difficult to segment because tumor regions have fuzzy boundaries and varying structures, including necrotic center, active tumor, and edema.
- Manual segmentation across increasing numbers of MRI scans is nearly impossible, motivating automated and semi-automated alternatives.
- Existing structured approaches can be computationally expensive, while machine-learning methods face data imbalance and limited clinical utility.
- The proposed patch-based nexus CNN uses dropout, batch normalization, inception-style processing, and two-phase weighted training to exploit local and contextual information efficiently.
2. Proposed Methodology
The proposed methodology segments brain tumors through a three-stage pipeline comprising pre-processing, CNN prediction, and post-processing. Pre-processed image patches are passed through a convolutional neural network to predict labels for individual patches.
- The approach evaluates convolutional neural networks for the brain tumor segmentation task.
- The methodology comprises pre-processing, CNN prediction, and post-processing.These three steps are presented as the overall segmentation pipeline.
- Pre-processed input images are divided into patches before CNN analysis.
- The convolutional neural network predicts output labels for individual patches.
2.1. Pre-processing
Pre-processing corrects MRI intensity artifacts and normalizes images before patch extraction. The resulting co-centric patches provide inputs for segmentation while correction improves contrast near lesion edges.
- MRI motion and field-inhomogeneity artifacts can create false intensity levels and false positives in automated segmentation.
- N4ITK bias field correction is used to remove unwanted artifacts from MR images.The correction is applied using the 3D Slicer toolkit version 4.6.2.
- After correction, the MRI scan shows better contrast near the edges, reducing a source of false positives.
- Intensity normalization brings the mean toward zero and variance toward one, while trimming the top and bottom one percent of intensities.This places intensities in a coherent range across images to facilitate learning during training.
- Two co-centric patches are extracted from each slice: a larger M × M patch and a smaller m×m patch.The patches are used in the concatenated architecture and are normalized.
2.2. Convolutional Neural Networks
CNNs hierarchically transform MRI patches through convolution, pooling, fully connected, and output layers to predict pixel labels. The proposed nexus variants combine CNN pathways to model tumor-segmentation features and dependencies.
- CNN architecture: CNNs stack convolutional, pooling, and fully connected layers hierarchically to transform inputs into output probabilities.Convolutional layers generate feature maps, pooling retains relevant features, and fully connected layers compute class probabilities.
- CNN architecture: The proposed model uses four MRI modalities as input patches and predicts the output label for each patch’s central pixel.The modalities are T1, T1c, T2, and T2-Flair.
- CNN architecture: Softmax converts network outputs into class probabilities, and the class with the largest probability is assigned to the corresponding central pixel.The output layer uses softmax as its non-linearity.
- Nexus architectures: Nexus architectures concatenate network outputs with later inputs to form connected models that model dependencies between neighboring pixels.The linear nexus treats the first CNN’s output as an additional channel for the second CNN.
- Nexus architectures: Five nexus architectures combine CNN pathways or modules, including linear, two-path, TLinear, inception, and ILinear nexus designs.ILinear combines linear nexus with inception modules and is described as incorporating speed and precision.
2.3. Post-processing
Post-processing uses morphological opening and closing to reduce small false positives introduced near the skull and image edges.
- Morphological cleanup: Morphological opening and closing apply erosion and dilation successively to remove small false positives around segmented-image edges.The procedure addresses false positives associated with high intensity around the skull.
3. Experimental Setup
The experimental setup specifies that dataset details, implementation parameters, and training procedures are described in subsequent subsections.
- Experimental setup: The paper organizes its experimental setup around dataset details, implementation parameters, and the training procedure.These components are discussed in the following subsections.
3.1. Dataset
Experiments use BRATS 2013 and BRATS 2015 multimodal MRI datasets with labeled tumor classes. Evaluation groups these labels into enhancing, core, and complete tumor categories.
- Datasets: BRATS 2013 and BRATS 2015 provide four MRI modalities—T1, T1c, T2, and T2flair—with segmentation labels for training data.The datasets differ substantially in the number of training images.
- Datasets: BRATS 2013 contains 30 training MR images, while BRATS 2015 contains 274 training MR images.BRATS 2013 includes 20 HGG and 10 LGG images; BRATS 2015 includes 220 HGG and 54 LGG images.
- Labels and evaluation: Training labels cover necrosis, edema, non-enhancing tumor, enhancing tumor, and healthy tissue, while evaluation uses three tumor groupings.The groupings are enhancing tumor, core tumor, and complete tumor.
3.2. Neural Network Parameters
The network parameters combine activation choices, batch normalization, dropout regularization, optimization settings, and implementation details to control learning and over-fitting.
- Implementation: The implementation uses Keras with TensorFlow or Theano back-ends, grid-search hyperparameter tuning, random weight initialization, and mostly zero biases.The final softmax-layer bias is set to 0.2, while convolution and max-pooling layers use stride 1.
- Activation functions: ReLU is selected for all proposed networks after max-out, ReLU, leakyReLU, and tangent activations are analyzed.Max-out and ReLU produce validation losses of 0.45 and 0.4, respectively, while leakyReLU and tangent cause under-fitting.
- Normalization: Batch normalization follows activation processing to keep activation statistics near zero mean and unit standard deviation.The normalized feature map is defined using batch normalization on weight parameter W_a followed by the ReLU non-linearity.
- Regularization: Dropout randomly removes feature-map activations during training to reduce over-fitting by encouraging units to learn independently.Dropout values are 0.5 in the first network half, 0.4 in the second half, and 0.3 before the final layer.
- Optimization: Stochastic gradient descent with the stated loss function updates network parameters during back-propagation.The loss uses target and predicted class labels over mini-batches.
3.3. Training
Training addresses class imbalance and optimization through mini-batch gradient methods, Nesterov acceleration, and a two-phase sampling and weighting procedure.
- Optimization: Mini-batch training reduces memory and processing requirements by propagating gradients from each batch through the network.Gradients update weights across all layers and parameters after each mini-batch.
- Optimization: Nesterov accelerated gradient estimates gradient change before updating, making optimization more responsive than classical momentum.The method modifies the velocity update using the momentum coefficient and an additional term based on the previous velocity.
- Two-phase training: The two-phase procedure first gives labels equal representation, then restores their uneven distribution to address severe pixel-level class imbalance.Healthy pixels comprise 98% of image labels, while the remaining 2% are divided among tumor classes.
- Two-phase training: The first phase trains for 20 epochs on 200,000 examples with equal class representation.The second phase trains only the output layer for 5 epochs on 30,000 patches while keeping other layers fixed.
- Two-phase training: Second-phase class weights are assigned according to training distribution, including weights of 8 for normal patches and 2 for edema.Core-tumor classes receive weight 1, preserving the true distribution while equal numbers of class patches are used.
- Architecture and settings: The first CNN receives a 33×33 patch and produces a 15×15 output matching the second CNN input dimensions.The network uses momentum coefficient µ = 0.9 and a learning rate decreased from lr = 0.01 to lr = 0.01 × 10^-4.
3.4. Evaluation Parameters
Evaluation measures segmentation quality with Dice similarity coefficient, sensitivity, and specificity across complete, core, and enhancing tumor classes.
- Metrics: The proposed segmentation is evaluated using Dice similarity coefficient, sensitivity, and specificity.These metrics are computed for complete tumor, core tumor, and enhancing tumor classes.
- Metrics: Dice similarity coefficient measures overlap between the predicted output image L and manually segmented label G.It is computed from the pixel-by-pixel intersection of the two images.
- Metrics: Sensitivity measures how accurately tumor labels are classified and how well the model detects tumor in an image.The measure focuses on correctly classified tumor labels.
- Metrics: Specificity measures correct classification of the normal output class by intersecting predicted normal labels with actual normal labels.It indicates how well the model confines tumor predictions to labeled tumor regions.
4. Experimental Results and Discussion
Experiments on BRATS 2013 and BRATS 2015 evaluate the proposed architectures, training strategy, and computational behavior. The results favor ILinear and two-phase training, especially for core and enhancing tumor regions, while identifying weaker edema-feature learning.
- Regularization results: Dropout after each convolution layer provides good validation accuracy, whereas dropout only before the output layer causes over-fitting and poor testing accuracy.The misleading validation effect is significantly evident in BRATS 2015, where fewer training patches limit feature learning.
- Training results: Two-phase training improves performance and removes most false positives compared with single-phase training.Single-phase models distribute labels more evenly and underperform across several parameters, whereas two-phase models follow the overall label distribution.
- Architecture comparison: ILinear with three parallel layers achieves the optimal performance after segmentation improvements plateau beyond four linear layers or three parallel layers.The evaluated architectures progress from linear and TLinear models to two-path, ILinear, and inception nexus designs.
- Segmentation quality: The proposed architectures achieve high specificity and strong results in core and enhancing regions, indicating accurate true-negative classification.The reported specificity values are associated with limited false positives and accurate tumor localization.
- Architecture comparison: ILinear stands out across performance metrics by combining strengths of inception nexus and linear architectures.Inception nexus performs better on core and enhancing metrics, while the linear model performs better on complete tumor; ILinear combines these behaviors.
- Benchmark comparison: On BRATS benchmarks, the proposed methods outperform state-of-the-art techniques in core and enhancing regions while producing comparable whole-tumor results.The paper identifies improved core and enhancing performance as a central result, with high specificity values.
- Computational efficiency: The proposed algorithm segments a whole brain in 5−10 minutes using a CPU, compared with up to 100 minutes for the best architecture reported in the benchmark literature.The paper presents this comparison as evidence of computational efficiency.
- BRATS 2015: The ILinear nexus achieves the best BRATS 2015 performance by learning features of all classes effectively.The proposed architectures are particularly effective at specifying tumor to its substantial area, generating high specificity values.
5. Conclusion
The proposed ILinear nexus architecture improves brain-tumor segmentation performance on BRATS 2013 and BRATS 2015 compared with state-of-the-art techniques. Its parallel-plus-linear design and weighted training support performance across both datasets, while larger datasets and more computational resources may enable further improvement.
- Automated brain-tumor segmentation with deep convolutional neural networks was evaluated on BRATS 2013 and BRATS 2015.
- The proposed architecture improves segmentation performance compared with state-of-the-art techniques on both datasets.
- The ILinear nexus stacks two networks: one with parallel layer placement and another with linearly arranged layers.
- Two-phase weighted training helps the network adapt to class imbalance in the data.
- The architecture performs particularly well for core and enhancing regions and shows high specificity in identifying substantial tumor areas.
- Performance could improve further with more training examples and computational resources.