Source-linked AI summary
DeepSeg: Deep Neural Network Framework for Automatic Brain Tumor Segmentation using Magnetic Resonance FLAIR Images
Ramy A. Zeineldin, Mohamed E. Karar, Jan Coburger, Christian R. Wirtz, Oliver Burgert
TL;DR
Brain tumor boundaries are difficult to distinguish from healthy tissue, motivating automated segmentation for clinical use. The paper proposes DeepSeg, a modular modified U-Net framework that combines interchangeable encoders and decoders for FLAIR MRI segmentation. Its architectures achieved validation mean DSC scores from 0.809 to 0.839, while the study notes scope limitations from using FLAIR data and ongoing expansion to other modalities and 3-D processing.
Problem
Distinguishing tumor tissue from normal brain parenchyma is difficult, while manual segmentation requires multiple experts and labor-intensive slice-by-slice decisions.
Method
DeepSeg is a modular decoupled framework based on modified U-Net, combining feature-extraction encoders with image-upscaling decoders and supporting different architectures.
Results
0.809 to 0.839 mean DSC scores were obtained on the validation set, with DenseNet achieving a DSC score of 0.839.
Takeaways & Limitations
The study demonstrated the feasibility of fully automated brain tumor segmentation in MR FLAIR images using different deep CNN architectures.
Takeaways & Limitations
The methods were trained using only FLAIR MRI data, and the framework’s validation is being extended with other modalities and 3-D convolutions.
Abstract
from arXiv · showhide
Purpose: Gliomas are the most common and aggressive type of brain tumors due to their infiltrative nature and rapid progression. The process of distinguishing tumor boundaries from healthy cells is still a challenging task in the clinical routine. Fluid-Attenuated Inversion Recovery (FLAIR) MRI modality can provide the physician with information about tumor infiltration. Therefore, this paper proposes a new generic deep learning architecture; namely DeepSeg for fully automated detection and segmentation of the brain lesion using FLAIR MRI data. Methods: The developed DeepSeg is a modular decoupling framework. It consists of two connected core parts based on an encoding and decoding relationship. The encoder part is a convolutional neural network (CNN) responsible for spatial information extraction. The resulting semantic map is inserted into the decoder part to get the full resolution probability map. Based on modified U-Net architecture, different CNN models such as Residual Neural Network (ResNet), Dense Convolutional Network (DenseNet), and NASNet have been utilized in this study. Results: The proposed deep learning architectures have been successfully tested and evaluated on-line based on MRI datasets of Brain Tumor Segmentation (BraTS 2019) challenge, including s336 cases as training data and 125 cases for validation data. The dice and Hausdorff distance scores of obtained segmentation results are about 0.81 to 0.84 and 9.8 to 19.7 correspondingly. Conclusion: This study showed successful feasibility and comparative performance of applying different deep learning models in a new DeepSeg framework for automated brain tumor segmentation in FLAIR MR images. The proposed DeepSeg is open-source and freely available at https://github.com/razeineldin/DeepSeg/.
Abstract
The paper concerns brain tumor, computer-aided diagnosis, convolutional neural networks, and deep learning.
- Brain tumor segmentation is the paper’s application domain.
- Computer-aided diagnosis is a related application context.
- Convolutional neural networks and deep learning are the paper’s machine-learning focus.
Introduction
Brain tumor surgery and diagnosis require reliable visualization and segmentation because tumor tissue can be difficult to distinguish from healthy brain tissue. The paper motivates modified U-Net architectures and a modular DeepSeg design for automated MRI segmentation using encoder–decoder paths.
- Introduction: Gliomas invade surrounding tissue, and high-grade glioma or glioblastoma has a median survival rate of one to two years.
- Introduction: Distinguishing tumor tissue from normal brain parenchyma is difficult during neurosurgical visual inspection.
- Introduction: MRI provides information about tumor cellularity, vascularity, and the blood-brain barrier through multimodal protocols.
- Introduction: Manual tumor segmentation requires multiple clinical experts and is labor-intensive and dependent on physicians’ experience and slice-by-slice decisions.
- Introduction: Classical automated segmentation approaches depend heavily on application-specific features and do not generalize well across tumors varying in shape, size, and localization.
- Introduction: Modified U-Net semantic segmentation uses contracting encoder and expansive decoder paths to combine semantic and spatial information.
- Introduction: The study proposes DeepSeg as a modular design for FLAIR MRI, with feature extraction and image-expanding paths supporting different deep neural network models.
Methods
DeepSeg is a modular, decoupled segmentation framework that combines interchangeable CNN feature extractors with a modified U-Net decoder for FLAIR MRI tumor segmentation. The method uses multimodal clinical data, preprocessing, augmentation, and several encoder designs to recover full-resolution tumor maps.
- DeepSeg framework: DeepSeg separates segmentation into an encoder for feature extraction and a decoder for image upscaling, enabling interchangeable model components.The framework is based on U-Net and supports straightforward comparison of different encoders and decoders.
- Feature extractors: The benchmarking system integrates modified U-Net encoders with dropout, batch normalization, and state-of-the-art CNN feature extractors.VGGNet serves as a simple baseline, while ResNet, DenseNet, Xception, MobileNet, MobileNetV2, and NASNet provide alternative feature-extraction designs.
- Feature extractors: ResNet uses shortcut connections in residual blocks to support training very deep networks without adding parameters or computational complexity.The paper describes training up to 152 layers with lower complexity than the VGG models.
- Decoder: The decoder recombines bottleneck semantic features with higher-resolution encoder outputs through skip connections to recover a full-resolution segmentation map.The modified decoder inserts batch normalization between convolution and ReLU, uses 32 base filters that double across layers, and processes full-size inputs rather than patches.
- Data: The study evaluates FLAIR MRI from BraTS 2019, whose 336 glioma patients have heterogeneous scans acquired under varied protocols and scanners across 19 institutions.The dataset includes T1-weighted, postcontrast T1-weighted, T2-weighted, and T2-FLAIR scans, with expert manual segmentations resampled to 1 mm3 resolution.
- Preprocessing and augmentation: Preprocessing applies improved N3 bias correction and slice-wise FLAIR normalization, while augmentation uses flips, rotation, scaling, shearing, shifts, and elastic distortion.Elastic distortion is described using elasticity coefficient σ and displacement multiplier α; horizontal and vertical flips are each applied to 20% of images.
Experiments
DeepSeg experiments evaluated multiple encoder variants with FLAIR MRI on BraTS 2019, combining segmentation accuracy, boundary distance, and computational measurements. Performance varied across encoders, with DenseNet and Xception achieving the highest DSC on training data, while MobileNet minimized Hausdorff distance and resource use.
- Experimental setup: The experiments used FLAIR MRI images at 224x224 resolution, training all networks for 35 epochs with batch size 16.Spatial dropout with rate 0.5 was also applied during training.
- Computational performance: 22 MB memory and roughly 5.6 thousand parameters made MobileNet the smallest-resource encoder, whereas deeper models required longer training.U-Net required 381 seconds per epoch and 1.1 seconds for prediction, while NASNet required 684 seconds per epoch and 4.4 seconds per patient; MobileNet variants shared 385-second training times because of smartphone-oriented internal architectures.
- Experimental setup: Two-fold cross-validation divided 336 BraTS 2019 training cases into 270 training and 66 validation cases.A separate BraTS 2019 validation evaluation used 125 cases.
- Segmentation performance: 0.839 DSC was achieved by both DenseNet and Xception, the highest scores among the tested encoders on the training-data evaluation.VGGNet and MobileNet followed with DSC scores of 0.837 and 0.835, respectively.
- Segmentation performance: 10.924 HD was obtained by MobileNet, the shortest Hausdorff distance among the reported training-data encoder results.Xception had the highest sensitivity at 0.856, approximately 7% better than the original U-Net model, while specificity was 0.998 for the listed encoders.
- Validation evaluation: 0.809 to 0.839 was the range of mean DSC scores on the validation set, compared with about 0.85 for expert whole-tumor-core annotation.The reported validation-set table also included sensitivity, specificity, and Hausdorff distance for the tested models.
Conclusions
DeepSeg demonstrated feasible automated brain-tumor segmentation in FLAIR MRI using different deep CNN architectures, with findings validated through the BraTS online platform. Further validation across MRI modalities and three-dimensional processing is planned to address clinical requirements.
- DeepSeg successfully enabled fully automated brain-tumor segmentation in MR FLAIR images using different deep CNN architectures.
- The comparative findings were validated using the BraTS online evaluation platform.
- Future work will extend validation with T1- and T2-weighted MRI datasets to examine potential effects on brain-tumor surgery planning.
- The framework is being advanced with three-dimensional convolutions and atrous spatial pyramid pooling to cover clinical requirements for accurate segmentation during MRI-guided interventions.
- The authors report DAAD scholarship funding, no conflicts of interest, no human or animal studies, and no patient data.