Medical Imaging Classification with Cold-Atom Reservoir Computing using Auto-Encoders and Surrogate-Driven Training

arXiv:2605.06727 · cs.LG, cs.ET, eess.IV · Submitted 2026-05-07 · Read on arXiv

Listen

Transcript

Introduction to the show: ident: AI Radio. Generated commentary on the latest Artificial Intelligence papers.

Tom: Today's paper: "Medical Imaging Classification with Cold-Atom Reservoir Computing using Auto-Encoders and Surrogate-Driven Training".

Jane: The gist: We introduce a hybrid quantum-classical pipeline, based on neutral-atom reservoir computing, for medical image classification, focusing on the binary classification task of polyp detection.

Tom: First, who's behind it and why it matters.

Paper summary: Jane: Basically, they are tackling this high dimensionality problem by first using an autoencoder to get a compact representation of the image data. Tom Right, so the autoencoder squashes the big image into something smaller, a latent space z. Lu And then they use that latent space to drive their quantum part.

Jane: But here's where it gets interesting: they have a problem with quantum measurements being non-differentiable for training. Tom That’s the 'gradient barrier'. So instead of trying to train directly on the quantum layer, they introduce a surrogate model.

Lu: They build this differentiable surrogate model that mimics how the quantum layer would behave, which lets them do end-to-end backpropagation through everything. Jane And they jointly optimize two things: getting a good classification accuracy and making sure the autoencoder still reconstructs the image accurately.

Meng: From a deployment standpoint, that joint optimization is smart because you don't want one part of the system to ruin the other's performance when you push it into real use. Tom It sounds like they learned a way to train this hybrid setup without getting stuck on those nasty quantum measurement problems.

Lalam: Based on what I’ve read, their results show that this approach outperforms traditional methods like PCA or unguided autoencoders when used as preprocessing steps for the reservoir computing pipeline. Jane So, when you look at the embeddings they get from this system, they form tighter and more distinct clusters.

Tom: That makes sense; if the features are better separated, the final linear classifier should perform better on those polyp detections. Lu The paper mentions that their QGARS pipeline is designed to learn compact and discriminative representations of image data that are also well-suited for quantum reservoir computing.

Jane: So, it's not just about the quantum part; the autoencoder is doing heavy lifting in finding features that make sense for the physics of neutral atoms. Tom And they show that even with NISQ devices, this method is viable for real-world medical tasks.

Meng: I wonder how much classical compute overhead this surrogate modeling adds to the training time when you're trying to scale it up. Jane That’s a valid concern because if the classical part gets too slow, the quantum speed advantage disappears.

Lalam: The paper notes that they conduct ablation studies on various parameters, showing how important things like lambda, which controls that trade-off between reconstruction loss and classification loss, are in tuning performance. Tom So even though it’s a hybrid system, there's still a lot of classical tuning to get the best results.

Lu: And they found an optimal guidance weight around lambda equals zero point five, which seems like the sweet spot for balancing reconstruction fidelity and classification accuracy.

Jane: It shows that this isn't just some arbitrary setup; there’s a specific configuration that works better than others. Tom So, they are giving us a concrete starting point if we want to try this on our own data.

Conclusion: Tom: So, looking at the whole "Medical Imaging Classification with Cold-Atom Reservoir Computing using Auto-Encoders and Surrogate-Driven Training" paper, the main thing is that they built a system that lets us use quantum techniques for medical image tasks without getting bogged down by the usual training hurdles. Jane They are combining classical autoencoders with neutral-atom reservoir computing to solve the problem of classifying things like polyps in scans.

Lu: The authors show that this hybrid pipeline creates latent representations that are both compressed and really good at separating different types of images for the quantum part to work on. Tom It’s about creating a representation that is optimized for both data compression and quantum feature extraction simultaneously.

Jane: What does this mean for us, if we only listen to the show? It means that even though we are still in the noisy intermediate-scale quantum era, there’s a structured way to use these specific atoms to find patterns in medical scans that outperforms older classical methods like standard autoencoders. Tom They aren't just proving quantum computing is cool; they are showing a practical path for applying it where it can actually help with classification.

Meng: From an engineering side, the implication is that we can build more robust pipelines for real clinical data, provided we manage the classical overhead from those surrogate models effectively. Jane So, the authors are pointing toward a scalable architecture where you take a classical feature extractor and feed it into a quantum processor for the final decision-making step.

Lalam: I'm confident in what they've done; their results on synthetic polyp, CVC-ClinicDB, and binary MNIST datasets show this method is superior to the PCA- and autoencoder-based baselines they compared against. Tom So it’s not just theoretical promise; there are empirical tests showing it works better than what we typically use as a baseline.

Jane: It really comes down to how effectively you guide that training process with these surrogate models, which the paper shows is crucial for getting good results in this hybrid setup. Lu The paper points toward using these specific quantum embeddings derived from Rydberg Hamiltonian detuning parameters as inputs for a linear classifier.

Tom: So the main thing to remember is that this isn't a magic solution; it’s a carefully constructed pipeline where classical and quantum components work together in a very specific way to find better features for medical diagnosis.

Nuno Batista, Ana Morgado, Oscar Ferraz, Sagar Silva Pratapsi, Jorge Lobo, Gabriel Falcao

Instituto de Telecomunicaçoes, Department of Electrical and Computer Engineering, University of Coimbra, Portugal · ISR - Institute of Systems and Robotics, Department of Electrical and Computer Engineering, University of Coimbra, Portugal · CFisUC, Department of Physics, University of Coimbra

cs.LG, cs.ET, eess.IV

Submitted: 2026-05-07

Updated: 2026-10-14

Comments: 8 pages, 6 figures. Accepted to the 2025 IEEE International Conference on Quantum AI (IEEE QAI). Supported by FCT and the Open Quantum Institute (OQI). v2: corrects the stated Rabi frequency to $Ω= 2π$ rad/$μ$s

Journal ref: 2025 IEEE International Conference on Quantum Artificial Intelligence (QAI)

DOI: 10.1109/QAI63978.2025.00064

License: http://arxiv.org/licenses/nonexclusive-distrib/1.0/

Importance score: 70/100

The gist: The gist: We introduce a hybrid quantum-classical pipeline, based on neutral-atom reservoir computing, for medical image classification, focusing on the binary classification task of polyp detection.

Key concepts

Quantum Reservoir Computing (QRC): This technique uses the dynamics of many-body quantum systems, like neutral atoms, to process input data. Data is encoded into the system's parameters, and its evolution generates complex quantum states whose measurements provide rich features suitable for classification.
Guided Autoencoder: A classical autoencoder is used first to compress high-dimensional medical image data into a compact latent representation. This latent space is then guided by the quantum reservoir layer, ensuring the learned features are well-suited for the subsequent quantum processing step.
Surrogate Model: Because standard training methods cannot handle the non-differentiable nature of quantum measurements, a classical neural network (the surrogate model) is trained to mimic the output of the quantum layer. This allows for end-to-end backpropagation through the entire hybrid system.
Guided Training Loss: The system is trained by minimizing a composite loss function that balances two goals: accurate image reconstruction and correct classification of the polyp. A weighting parameter ($\lambda$) controls how much emphasis is placed on each objective during training.

Terminology

Embeddings — 2× in this paper · explained in 2 episodes
Ablation — 1× in this paper · explained in 1 episode
Ablation Studies — 1× in this paper · explained in 1 episode
Accuracy — 1× in this paper · explained in 1 episode
Latent Representation — 1× in this paper · explained in 2 episodes
Latent Representations — 1× in this paper · explained in 3 episodes
Robustness — 1× in this paper · explained in 5 episodes
Surrogate Model — 1× in this paper · explained in 3 episodes

Summary

The gist: We introduce a hybrid quantum-classical pipeline, based on neutral-atom reservoir computing, for medical image classification, focusing on the binary classification task of polyp detection.

How it works

The proposed pipeline is a Quantum-Guided Autoencoder with a Reservoir Surrogate (QGARS), which is an hybrid architecture that synergizes classical autoencoding techniques with quantum reservoir processing to achieve efficient learning of features The system is designed to overcome the inherent nondifferentiability of quantum operations, which traditionally impede gradient-based training methods.

The architecture consists of several key components:

** Classical Autoencoder:**

The encoder transforms high-dimensional input data x into a lower-dimensional latent representation z The decoder network Dρ(z) reconstructs the original input data from this latent representation, producing xˆ.

Quantum Reservoir Layer:

The system maps the latent representation z to local detuning frequencies ∆i of the Rydberg atoms The system evolves under a Hamiltonian H(∆), resulting in a quantum state ψ(∆)⟩. Measurements of observables on this state yield embeddings q ∈ Rk, which serve as training targets for the surrogate model or as inputs for the classification task later on.

Surrogate Model:

A differentiable classical neural network fsur(z; θsur) is trained to approximate the mapping from z to q, effectively emulating the quantum layer’s behaviour.

Guided Autoencoder

The Guided Autoencoder is trained to minimize a composite loss function that balances reconstruction fidelity and classification accuracy The loss function is defined as L = (1 − λ) · LR + λ · LC.

Loss Components:

The reconstruction loss LR measures the mean squared error between the input x and its reconstruction xˆ.

The classification loss LC is computed as the cross-entropy between the true labels yi and the predicted probabilities yˆi, which are obtained through a linear mapping of the approximated embeddings q′.

Hyperparameter:

The parameter 0 < λ < 1 controls the trade-off between reconstruction and classification objectives.

Surrogate Modeling for Quantum Layers

To circumvent the gradient barrier, surrogate models are used to approximate the input-output behavior of quantum circuits The surrogate model fsur(z; θsur) is implemented as a feedforward neural network designed to learn the mapping q from the autoencoder’s latent space to the quantum embeddings q produced by the quantum reservoir.

Training Procedure:

The surrogate model is trained by minimizing the loss function Lsur = ∥q′ − q∥2.

Regular updates are performed periodically during training to ensure the surrogate model accurately reflects the quantum reservoir’s behavior. This integration establishes a continuous computational graph from the output layer back to the input layer, enabling gradient-based optimization techniques.

Quantum Reservoir Computing Layer

The neutral-atom reservoir leverages many-body quantum dynamics to encode data in a space whose dimension grows exponentially with atom number.

Data Encoding:

Input data is encoded into the quantum system by mapping the latent representations from the autoencoder to the detuning parameters of the Rydberg Hamiltonian.

Quantum Readout and Embedding Dimensionality:

The quantum reservoir’s output is obtained by probing, in successive time steps, the state of the atoms after the systems evolution up to a certain time t. The dimensionality of the QRC embeddings generated in our pipeline is a direct consequence of the number of qubits (N) employed in the neutral-atom chain, the specific quantum mechanical observables measured, and the number of discrete time steps (T) over which these measurements are collected every ∆t. If all three proposed observables are measured, the total dimensionality q of the final QRC embedding vector is O(T N3) for the considered observables.

Comparison with Classical Methods

The QGARS pipeline outperforms traditional feature reduction methods commonly used as preprocessing steps on QRC pipelines The results show that the QGARS pipeline is able to outperform traditional feature reduction methods commonly used as preprocessing steps on QRC pipelines. The QGARS embeddings exhibit the tightest and most distinct clusters, indicating superior discriminative structure.

Conclusion

QGARS is a hybrid quantum-classical architecture that integrates a classical autoencoder with a neutral-atom quantum reservoir and a differentiable surrogate model to enable end-to-end training Our experiments on synthetic polyp, CVC-ClinicDB, and binary MNIST datasets demonstrate that QGARS outperforms classical PCA- and autoencoder-based baselines that are commonly used for quantum reservoir computing pipelines. Ablation studies highlight the importance of guidance weight λ, surrogate update frequency, and reservoir parameters. Although challenges remain—particularly regarding high-dimensional embeddings, surrogate fidelity, and classical overhead—our proposed extensions and hardware considerations chart a clear path toward robust, scalable quantumenhanced medical imaging pipelines in the NISQ era.

The results from these experiments are discussed in the next section V. The effect of Guided Lambda on QRC Accuracy is plotted in Fig. 5. We find optimal performance around λ = 0.5; too little guidance (λ < 0.1) approaches a point where it is equivalent to a traditional AE, so it yields under-trained classifiers. A strange effect that we verified was that even using the value λ = 1, the performance of the classifier is still competitive.

Limitations and Future Work

While QGARS demonstrates strong performance, several limitations remain. Classical overhead from frequent surrogate retraining and batch quantum simulations introduces non-trivial classical compute overhead, potentially offsetting some of the quantum advantage. Future work plans include exploring different Hamiltonians, replacing fully connected autoencoders with convolutional AE, investigating alternative surrogate architectures like PINNs, and running experiments on real quantum processing units (QPUs) with more qubits.

We have presented Quantum-Guided Autoencoder with Reservoir Surrogate (QGARS), a hybrid quantum-classical architecture that integrates a classical autoencoder with a neutral-atom quantum reservoir and a differentiable surrogate model to enable end-to-end training. By jointly optimizing reconstruction and classification losses, QGARS learns compact and discriminative features tailored for quantum reservoir computing. Our experiments on synthetic polyp, CVC-ClinicDB, and binary MNIST datasets demonstrate that QGARS outperforms classical PCA- and autoencoder-based baselines that are commonly used for quantum reservoir computing pipelines. Ablation studies highlight the importance of guidance weight λ, surrogate update frequency, and reservoir parameters. Although challenges remain—particularly regarding high-dimensional embeddings, surrogate fidelity, and classical overhead—our proposed extensions and hardware considerations chart a clear path toward robust, scalable quantumenhanced medical imaging pipelines in the NISQ era.

The QGARS architecture is illustrated in Fig. 2. The total number of parameters for the linear classifier with a reservoir containing 12 qubits is 9538. The QGARS embeddings exhibit the tightest and most distinct clusters, indicating superior discriminative structure.

Improvements for AI systems

textbf Surrogate Model Fidelity Enhancement for Quantum Layers: Replace standard feedforward networks with Physics-Informed Neural Networks (PINNs) as suggested in Section VI B, Potential Extensions. This would allow the surrogate model to better capture the complex, non-linear transformations inherent in quantum dynamics by incorporating physical laws directly into the loss function, potentially improving model fidelity beyond current approximations like Eq. (11).
textbf Context-Aware Quantum Encoding: Adapt the data encoding scheme from Section E.1 to incorporate learned context from a classical component. Instead of mapping latent representations directly to detuning parameters, develop a mechanism where the autoencoder's latent space informs the choice or structure of the Hamiltonian H(t), allowing the quantum reservoir dynamics to be dynamically tailored to specific input feature subsets.
textbf Scalable Quantum Reservoir Dynamics: Investigate alternative interaction Hamiltonians beyond Eq. (5), such as those from the Heisenberg XXZ model mentioned in Section VI B, Potential Extensions. This exploration could unlock richer patterns in medical data by allowing the reservoir to explore different types of many-body correlations that are more relevant to high-dimensional image classification tasks.
textbf Adaptive Training Schedule via Surrogate Update Frequency: Systematically optimize the Quantum Update Frequency hyperparameter mentioned in Section F, Parameter Sweep Strategy. This optimization should be performed using Bayesian methods to automatically determine the optimal frequency that balances minimizing classical overhead with maximizing surrogate model accuracy and classification performance, moving beyond fixed epoch sweeps.
textbf Spatially Aware Feature Extraction: Replace the classical autoencoder described in Section III A with a Convolutional Autoencoder (CAE) as suggested in Section VI B. This would allow the encoder to learn spatially localized features from medical images, providing a richer structure for the subsequent quantum reservoir layer to process, thereby improving the representation of image textures and shapes.