SegDGAN

SegDGAN performs automated prostate gland segmentation from magnetic resonance imaging (MRI) using generative adversarial networks (GANs) to improve delineation accuracy for diagnosis and treatment planning.

Key Features:

• Automatic Segmentation: Produces automatic prostate gland delineations from MRI to replace manual segmentation steps.
• Fully Convolutional Generator: Implements a fully convolutional generation network incorporating densely connected blocks to enhance feature learning and segmentation precision.
• Multi-Scale Critic Network: Employs a critic network configured for multi-scale feature extraction to improve adversarial supervision of segmentation outputs.
• Loss Functions: Optimizes an objective function combining mean absolute error (MAE) and Dice coefficient to align predictions with ground truth segmentation.
• Comparative Evaluation: Evaluated against U-Net, FCN, and SegAN on a 220-patient clinical dataset and the public PROMISE12 dataset using DSC, VOE, ASD, and HD metrics.
• Quantitative Performance: Reported results include DSC=91.66%, VOE=15.28%, ASD=0.51 mm, HD=11.58 mm on the clinical dataset, and DSC=86.24%, VOE=23.60%, ASD=1.02 mm, HD=7.57 mm on PROMISE12.

Scientific Applications:

• Prostate MRI segmentation for diagnosis: Provides delineations that support prostate disease detection and localization in MRI.
• Treatment planning: Supplies segmentation outputs usable for planning interventions and radiotherapy targeting.
• Monitoring and longitudinal studies: Enables consistent segmentation across datasets to support follow-up and comparative analyses.

Methodology:

Uses a classic GAN framework with a fully convolutional generator containing densely connected blocks and a critic network for multi-scale feature extraction, trained by optimizing mean absolute error and Dice coefficient.