I-Translation: CT ↔ MRI Synthesis
A novel Quad-(Cycle)GAN architecture of 8 neural networks (4 generators + 4 discriminators) performing bidirectional unpaired brain CT↔T2 MRI translation — trained on 145 images from 20 patients, achieving FID 20 with 7,000+ epoch stability.
Quad-(Cycle)GAN Architecture — 8 Neural Networks
The architecture doubles CycleGAN's 4-network design to 8 networks — creating 4 independent cycle paths and a self-correcting gradient feedback system. This eliminates mode collapse and enables training beyond 7,000 epochs, versus CycleGAN which collapses between epochs 500–2,000.
Base architecture: U-Net with Instance Normalisation · 40 layers per generator · 4×4 kernels throughout · PatchGAN discriminators (70×70 receptive field, 8×8 output) · λ_cyc=10, λ_id=0.5, λ_adv=1
4 Generators — U-Net Based, Instance Normalised
All 4 generators share an identical U-Net architecture (6-layer encoder → bottleneck → 6-layer decoder with 5 skip connections) but carry distinct trained weights (~50 MB each), resulting in different learned feature representations. Generators G and F produce the representative outputs — their conversions are shown as demonstration results below.
4 Discriminators — PatchGAN (70×70 Receptive Field · 8×8 Output)
U-Net Generator Architecture — Per Generator (×4)
Loss Function — Total Objective
Ltotal = λadv·Ladv_primary + λcyc·Lcyc_primary + λadv·Ladv_aux + λcyc·Lcyc_aux + λid·Lidentity
Structural preservation priority
Domain realism enforcement
Domain consistency preservation
Fixed loss weights — empirically tuned. Constant learning rate: 0.0002. Batch size: 8–16. Framework: TensorFlow 2.4.
4 Independent Cycle Paths — The Core Innovation
Proprietary 12-Stage Deterministic Denoising Pipeline
Raw GAN output undergoes twelve precision-engineered post-processing stages to achieve clinical-grade image quality.
Validated outcomes: 68% noise reduction · 53% brightness consistency improvement · Cohen's d = 1.23 · p < 0.01
App Demo — Bidirectional CT ↔ MRI Conversion
Real-time output from the I-Translation web application demonstrating both conversion directions via the representative generators G and F.
Demo outcome only. Demonstrates conversion capability — not clinical validation.
Demo outcome only. Demonstrates conversion capability — not clinical validation.
Performance vs State of the Art
| Metric | CycleGAN (Baseline) | Quad-(Cycle)GAN | Improvement |
|---|---|---|---|
| FID Score ↓ | 61.4 ± 3.7 | 45.8 ± 2.5 | 25.4% · p<0.001 · d=5.12 |
| KID Score ↓ | 0.048 ± 0.005 | 0.028 ± 0.003 | 41.7% · p<0.001 · d=4.89 |
| LPIPS Score ↓ | 0.302 ± 0.015 | 0.252 ± 0.010 | 16.6% · p<0.001 · d=4.12 |
| IoU Score ↑ | 0.59 ± 0.08 | 0.71 ± 0.05 ✓ | 20.3% · above clinical threshold |
| Gradient Variance ↓ | 0.158 | 0.041 | 3.85× reduction |
| Max Stable Epochs | ~500–2,000 | 7,000+ (no collapse) | 3.5–7× longer |
| Neural Networks | 4 (2G + 2D) | 8 (4G + 4D) | 2× · 4 cycle paths |
Statistical significance: all metrics p<0.001, Cohen's d >4 (huge effect size). Training: 5,000 epochs on 145 brain CT/MRI images from 20 patients.
QuadGAN — 10 Unique Architectural Characteristics
Source: Patent Application No. 202611033808 · COMPLETE TECHNICAL PATENT DOCUMENTATION
| # | Characteristic | Innovation / Specification | Clinical Benefit |
|---|---|---|---|
| 1 | Quad-Generator Diversity | 4 generators with same architecture, different weights | Multiple output options for radiologist selection |
| 2 | Instance Normalization | Normalize each patient independently | Better generalization to unseen patients |
| 3 | PatchGAN Discriminator | 70×70 receptive field, 8×8 output grid | High-frequency detail preservation |
| 4 | Checkpoint 652 | Intentionally under-trained (13% of full training) | Faster training, higher diversity, reduced overfitting |
| 5 | Fixed Loss Weights | λcyc=10, λid=0.5, λadv=1 (constant throughout training) | Stable training, prioritizes structure preservation |
| 6 | Symmetric U-Net | 6 encoder + 6 decoder layers with skip connections (40 total layers) | Gradient flow, feature preservation, spatial precision |
| 7 | Asymmetric Activations | LeakyReLU (encoder), ReLU (decoder), Tanh (output) | Robust training, prevents dying ReLU problem |
| 8 | Strategic Dropout | 50% dropout in first 3 decoder layers only | Regularization without sacrificing output detail |
| 9 | 64×64 Resolution | Ultra-low resolution for fast inference | <1s inference on CPU; suitable for research prototype |
| 10 | Lanczos Resampling | High-quality downsampling for preprocessing | Edge preservation, noise reduction in input images |
QuadGAN vs CycleGAN — Key Differentiators
| Dimension | QuadGAN | CycleGAN |
|---|---|---|
| Neural Networks | 8 (4G + 4D) | 4 (2G + 2D) |
| Cycle Paths | 4 | 2 |
| Output Diversity | 4 independent outputs | 1 output |
| Gradient Variance | 0.041 | 0.158 |
| Stable Training (epochs) | 7,000+ | ~3,000 |
| FID Score | 20 (overall) | 34+ (typical) |
All metrics sourced from patent documentation. Patent Application No. 202611033808.
Clinical Applications
Neurological Imaging
Brain tumour screening, stroke assessment, and dementia diagnosis from CT — without MRI access. Addresses India's 9M+ annual MRI gap.
Data Augmentation
Expand limited medical datasets with high-quality synthetic images for AI training, research, and privacy-preserving collaboration.
Oncology Staging
Tumour characterisation and treatment planning with synthesised multi-modal imaging at ₹500–₹3,000 vs ₹8,000–₹25,000 MRI scan cost.
Medical Education
Generate diverse training cases for medical education and radiology training, providing comprehensive cross-modal learning experiences.