3D Prostate MRI Segmentation Using GAN

Jan 31, 2025·
Syed Ibrar Hussain
· 2 min read
Abstract
This project presents a deep learning pipeline for automated 3D prostate segmentation from MRI (NIfTI) scans. It leverages a 3D U-Net-based generator with a PatchGAN discriminator to refine volumetric segmentations. The pipeline integrates preprocessing, data augmentation, GAN-based adversarial learning, and evaluation metrics for segmentation accuracy. The code repository is private and can be shared upon request. This approach improves reproducibility and demonstrates enhanced performance in prostate segmentation tasks for radiomics and clinical applications.
projects research project

Project Overview

This project focuses on automatic segmentation of prostate MRI scans using a Conditional Generative Adversarial Network (cGAN). The pipeline integrates volumetric 3D T2-weighted MRI data with advanced deep learning architectures to produce highly accurate slice-wise segmentations, later reconstructed into 3D volumes.

Key Features:

  • Preprocessing & normalization of 3D MRI volumes
  • Slice extraction for efficient training
  • U-Net based generator and PatchGAN discriminator
  • Adversarial training with combined L1 + GAN loss
  • K-Fold cross-validation for robust performance evaluation
  • Post-processing & morphological refinement for smooth and accurate segmentation
  • Performance metrics: Dice, Precision, Recall, Hausdorff distance

Repository: Private (Available upon request)

GAN Pipeline

3D Prostate GAN Pipeline

The illustration above summarizes the workflow from raw MRI input → preprocessing → GAN-based segmentation → 3D volume reconstruction → post-processing. This approach enables automated and reproducible segmentation, facilitating downstream radiomics analysis and clinical decision support.

Impact & Applications

  • Supports reproducible biomedical imaging research
  • Enables accurate prostate segmentation for radiomics and AI-driven diagnosis
  • Facilitates quantitative assessment of anatomical structures in 3D MRI
  • Demonstrates integration of generative AI models for clinical applications

Technical Details

The pipeline includes:

  1. Data Handling: 3D NIfTI MRI volumes are normalized and augmented.
  2. Model Architecture: 3D U-Net generator with PatchGAN discriminator for adversarial learning.
  3. Training Strategy: Combined L1 reconstruction loss + adversarial loss for realistic and accurate segmentations.
  4. Evaluation: Dice similarity, Precision, Recall, and Hausdorff distance metrics.
  5. Post-Processing: Morphological operations refine segmented masks for smooth 3D surfaces.

Note: The repository contains sensitive patient data and is private. Access can be granted upon request for research collaboration.

GAN 3D MRI Medical Imaging Prostate Segmentation
Authors
Syed Ibrar Hussain