Cross-Modality Medical Image Translation (MRI ↔ CT)

This repository contains an unsupervised deep learning framework for bidirectional medical image translation between MRI and CT modalities using CycleGAN. This project is specifically designed to generate high-fidelity Synthetic CT (sCT) from MRI scans to enable MRI-only radiotherapy planning, effectively reducing patient radiation exposure and clinical costs.

🚀 Key Features

• Unpaired Training: Operates without the need for perfectly aligned MRI-CT image pairs, utilizing cycle-consistency constraints.
• ResUNet Architecture: Features a ResNet-9 generator that preserves deep anatomical features and mitigates gradient vanishing.
• SSIM-Enhanced Training: Beyond standard $L_1$ and Adversarial losses, an explicit Structural Similarity (SSIM) Loss is integrated to ensure anatomical integrity.
• Medical Domain Optimized: Built using the MONAI (Medical Open Network for AI) framework for specialized medical data augmentation and intensity transforms.

🏗️ Technical Architecture

• Generators: ResNet-based models with 9 residual blocks and Instance Normalization.
• Discriminators: 70x70 PatchGAN with Spectral Normalization to ensure training stability and realistic local textures.
• Training Strategy: 60 epochs with a starting learning rate of 0.0002, utilizing a Cosine Annealing scheduler and the Adam optimizer.

📊 Results & Evaluation

The model demonstrates high fidelity in structural preservation. While the CT-to-MRI cycle shows higher stability, the MRI-to-CT translation successfully recovers skeletal structures necessary for dose calculation.

Metric	MRI Cycle ($MRI \rightarrow CT \rightarrow MRI$)	CT Cycle ($CT \rightarrow MRI \rightarrow CT$)
Mean SSIM	0.7660	0.9173
Peak SSIM	0.8511	0.9262
Mean PSNR	~24.5 dB	~29.8 dB

🛠️ Installation

Ensure you have the following dependencies installed:

pip install torch torchvision monai matplotlib numpy

📚 References

Yu, L., et al. (2025). "AI-enhanced PET/CT image synthesis using CycleGAN for improved ovarian cancer imaging," Polish Journal of Radiology, vol. 90.

Bahloul, M. A., et al. (2024). "Advancements in synthetic CT generation from MRI: A review of techniques," Frontiers in Radiology.

SciTePress (2024). "CT to MRI Image Translation Using CycleGAN: A Deep Learning Approach for Cross-Modality Medical Imaging."

Zhu, J.-Y., et al. (2017). "Unpaired Image-to-Image Translation using Cycle-Consistent Adversarial Networks," ICCV.

Development Team: Abhinab Das, Abhishek S U, Abhinav Pandey, Priyanshu Thakur