mediq-gan-medical-image-generation - SKILL.md Agent Skill

name: mediq-gan-medical-image-generation description: "Quantum-inspired GAN methodology for high-resolution medical image generation with prototype-guided skip connections and dual-stream generator. Addresses data scarcity, class imbalance, and privacy constraints in medical imaging through variational quantum circuits that preserve full-rank mappings and avoid rank collapse. Use when building quantum-inspired generative models for medical image augmentation, designing GAN architectures that balance expressivity with trainability, or analyzing latent geometry of quantum-inspired generators." metadata: arxiv_id: "2506.21015" date: "2025-06-26" authors: "Qingyue Jiao, Yongcan Tang, Jun Zhuang, Jason Cong, Yiyu Shi" tags: [quantum, gan, medical-imaging, data-augmentation, variational-quantum-circuit, generative-model]

MediQ-GAN: Quantum-Inspired GAN for Medical Image Generation

Core Concept

MediQ-GAN is a quantum-inspired Generative Adversarial Network for high-resolution medical image generation and data augmentation. It addresses the critical challenge that medical imaging datasets are often scarce, imbalanced, and constrained by privacy — making classical generative models inadequate due to their extensive computational and sample requirements.

The key innovation is a dual-stream generator that fuses classical and quantum-inspired branches, with prototype-guided skip connections. Its variational quantum circuits inherently preserve full-rank mappings, avoid rank collapse, and are theory-guided to balance expressivity with trainability.

Key Technical Insights

Dual-Stream Architecture: Classical branch handles spatial features; quantum-inspired branch captures high-dimensional correlations through variational quantum circuits (VQCs). The streams are fused via prototype-guided skip connections.
Rank Preservation: VQCs inherently maintain full-rank mappings, avoiding the rank collapse problem common in classical GANs. This is proven through latent-geometry and rank-based analysis.
Theory-Guided Expressivity-Trainability Balance: Unlike standard VQAs that face barren plateaus, MediQ-GAN's architecture is designed to balance circuit expressivity with gradient trainability.
Hardware-Agnostic Design: Validated on IBM quantum hardware for robustness, but the framework works with any quantum simulator or classical approximation of quantum circuits.

Implementation Patterns

Pattern 1: Dual-Stream Generator Design

class MediQGenerator(nn.Module):
    def __init__(self, latent_dim, n_qubits, n_layers):
        super().__init__()
        # Classical stream
        self.classical_branch = nn.Sequential(
            nn.Linear(latent_dim, 256),
            nn.ReLU(),
            nn.Linear(256, 128),
        )
        # Quantum-inspired stream
        self.quantum_branch = VariationalQuantumCircuit(
            n_qubits=n_qubits, n_layers=n_layers
        )
        # Prototype-guided skip connections
        self.skip_fusion = PrototypeGuidedFusion(dim=128)
        
    def forward(self, z):
        c_feat = self.classical_branch(z)
        q_feat = self.quantum_branch(z)
        return self.skip_fusion(c_feat, q_feat)

Pattern 2: Prototype-Guided Skip Connection

Prototype-guided skip connections use class prototypes (learned representative features) to modulate information flow between generator layers:

Learn class prototypes during training
At each skip connection, compute similarity between current features and prototypes
Weight the skip connection based on prototype similarity
This guides generation toward semantically meaningful outputs

Pattern 3: Latent Geometry Analysis

Validate quantum-inspired GAN quality through:

Rank analysis: Measure effective rank of feature covariance matrices
Latent geometry: Analyze manifold structure of generated vs. real samples
Full-rank preservation: Verify VQC maintains rank throughout training

Pattern 4: Data-Augmentation Pipeline

Train MediQ-GAN on limited medical dataset
Generate synthetic samples for minority classes
Use synthetic + real data for downstream classifier training
Evaluate on held-out real test set

Applications

Medical image data augmentation for rare diseases
Class imbalance mitigation in diagnostic datasets
Privacy-preserving synthetic medical data generation
Training data generation for downstream ML models
Quality assessment of quantum-inspired vs. classical generative models

Activation Keywords

quantum gan medical
mediq-gan
quantum-inspired image generation
medical image augmentation
dual-stream quantum generator
prototype-guided skip connection
variational quantum circuit gan
rank collapse prevention
quantum generative model
医疗图像生成
量子生成对抗网络
医学数据增强

Error Handling

Barren Plateaus in Quantum Branch

Reduce circuit depth or use layerwise training
Apply the expressivity-trainability analysis from hqnn-expressibility-trainability-nas skill

Mode Collapse

Increase prototype diversity in skip connections
Use minibatch discrimination in the discriminator
Apply spectral normalization

Hardware Simulation Overhead

Use classical shadow simulation for larger circuits
Reduce qubit count for initial experiments (4-8 qubits)
Validate on IBM hardware for final robustness check

Resources

Paper: https://arxiv.org/abs/2506.21015
Related: hqnn-expressibility-trainability-nas, quantum-generative-diffusion-medical, cold-atom-medical-imaging