DeepSeg

DeepSeg performs automated segmentation of brain tumors from FLAIR (Fluid-Attenuated Inversion Recovery) MRI images to delineate tumor boundaries for quantitative analysis and treatment planning.

Key Features:

• Architecture: A modular encoder–decoder architecture decouples feature extraction and reconstruction into separate components.
• Encoder: The encoder employs convolutional neural networks (CNNs) to extract spatial information and generate a semantic map from FLAIR MRI data.
• Decoder: The decoder reconstructs the semantic map into a full-resolution probability map for pixel-wise segmentation.
• Backbone: Builds upon a modified U-Net framework to support dense prediction at full resolution.
• CNN Variants: Incorporates ResNet, DenseNet, and NASNet architectures as CNN backbones.
• Input Modality: Operates on FLAIR (Fluid-Attenuated Inversion Recovery) MRI images.
• Target Pathology: Focuses on gliomas and infiltrative brain tumor regions.
• Evaluation Dataset: Evaluated on the Brain Tumor Segmentation (BraTS) 2019 dataset with 336 training cases and 125 validation cases.
• Performance Metrics: Reported Dice scores between 0.81 and 0.84 and Hausdorff distances between 9.8 and 19.7.

Scientific Applications:

• Brain tumor segmentation: Automated delineation of tumor regions in FLAIR MRI for volumetric and spatial analysis.
• Boundary delineation: Distinguishing tumor boundaries from healthy tissue to assess infiltrative tumor extent.
• Treatment planning: Providing probability maps and quantitative metrics to inform radiotherapy and surgical planning.
• Neuro-oncology research: Benchmarking segmentation methods and evaluating algorithm performance using BraTS data.

Methodology:

Modular encoder–decoder approach where CNN-based encoders generate semantic maps and decoders reconstruct full-resolution probability maps; implemented on a modified U-Net backbone with ResNet, DenseNet, and NASNet variants and evaluated on BraTS 2019 (336 training, 125 validation) using Dice and Hausdorff metrics.