💻 computer science

TE-HQViT-Probs: Full-Probability Quantum Self-Attention Outperforms Expectation-Value Readout in Medical Image Classification.

This study demonstrates that utilizing full-probability readout instead of expectation-value measurements in a compact hybrid quantum self-attention model can improve medical image classification performance on specific datasets, highlighting readout design as a critical architectural factor while noting that the findings do not yet establish quantum advantage or clinical readiness.

Original authors: Thanh Phong Nguyen

Published 2026-08-18

📖 4 min read☕ Coffee break read

CC BY 4.0

Original authors: Thanh Phong Nguyen

Original paper licensed under CC BY 4.0 (https://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

In the rapidly evolving field of artificial intelligence, computers are increasingly learning to see by breaking images into small, manageable pieces, much like a mosaic. These pieces, or tokens, are then analyzed to understand how they relate to one another, a process that allows machines to identify complex patterns in everything from satellite photos to medical scans. Recently, scientists have begun experimenting with a hybrid approach, mixing standard computer processors with quantum circuits—devices that use the strange rules of quantum physics to process information. The hope is that these quantum components can uncover hidden details in data that traditional computers might miss. However, a critical step in this process is how the quantum machine translates its internal, invisible state into a form the rest of the computer can understand. This translation, known as a "readout," is the bridge between the quantum world and the classical world. If this bridge is too narrow, it might discard valuable information before the computer can use it to make a diagnosis.

A researcher at Van Lang University in Vietnam set out to test whether the design of this bridge matters more than the quantum machinery itself. The study focused on a specific type of medical image analysis, using three different datasets containing pictures of pneumonia in lungs, breast tissue, and organ structures. The researcher built a model that acted like a hybrid vision system. In this system, the computer calculated some parts of the image analysis using standard methods, but it passed the most important data through a small quantum circuit containing just four quantum bits, or qubits. The core question was simple: would the model perform better if it measured the quantum circuit by looking at the full probability of every possible outcome, or if it used a simpler, more common method that only measured the average behavior of individual qubits?

The results showed that the more detailed measurement method, which captures the full probability distribution, consistently provided better results than the simpler average-based method. In simulations, the model using the full probability readout improved its ability to distinguish between different medical conditions. On one dataset involving breast tissue, the improvement was particularly noticeable, raising the model's accuracy score by nearly five percentage points compared to the simpler method. On another dataset focused on organ structures, the improvement was also significant. However, the gains were not uniform; on the pneumonia dataset, the difference was very small. This suggests that while capturing more information helps, the benefit depends heavily on the specific type of data being analyzed.

Crucially, the study found that this improvement did not come from the quantum circuit "learning" in the way a human student learns. The researcher tested this by freezing the quantum circuit's settings so they could not change during training, essentially turning off the quantum learning process. Surprisingly, the model still performed well when using the full probability readout, even with these frozen settings. This indicates that the advantage came from the way the information was measured and presented to the computer, rather than from the quantum circuit itself finding a clever new solution. The detailed measurement simply revealed information that the simpler method had thrown away.

Despite these findings, the study is careful not to claim that quantum computers have solved medical diagnosis or that they are faster than current technology. The experiments were run entirely on classical computers simulating quantum behavior, using only four qubits, which is a very small number in the quantum world. The models tested were also compared against standard, non-quantum artificial intelligence systems, and in several cases, the traditional systems performed just as well or better. The research concludes that the design of the measurement process is a vital, often overlooked part of building hybrid quantum systems. It suggests that before scientists can claim a true quantum advantage, they must first ensure they are not accidentally discarding useful data through a poorly designed readout. For now, the most important takeaway is that in the quest to combine quantum physics with artificial intelligence, how you look at the result matters just as much as the quantum machine itself.

Technical Summary: TE-HQViT-Probs

Problem Statement
Hybrid quantum-classical neural networks typically convert parameterized quantum circuit (PQC) states into classical features using expectation-value measurements (e.g., Pauli-Z observables). While compact, this readout method summarizes the quantum state through a limited set of observables, potentially discarding class-relevant information contained in the full quantum state. The central research question is whether utilizing a full computational-basis probability readout—exposing the probability distribution over all measured basis states—provides superior classification features compared to standard expectation-value readout within a compact hybrid quantum self-attention architecture. The study explicitly aims to avoid claims of quantum computational advantage or hardware speedup, focusing instead on architectural design and information preservation in small-qubit simulations.

Methodology
The author proposes TE-HQViT-Probs (Token-Embedded Hybrid Quantum Vision Transformer with full-probability readout), a hybrid model applied to medical image classification tasks using the MedMNIST benchmark (PneumoniaMNIST, BreastMNIST, and OrganAMNIST).

Architecture: The model follows a Vision Transformer (ViT) structure where images are divided into patch tokens.
- Query and Key Projections: Computed entirely via classical linear layers.
- Value Path: Processes token embeddings using a hybrid quantum module. This involves:
  1. Amplitude Embedding: Mapping the input vector to a quantum state $|\psi_{in}\rangle$ .
  2. Parameterized Quantum Circuit (PQC): A 4-qubit, 2-layer variational circuit with trainable single-qubit rotations and nearest-neighbor entangling gates.
  3. Readout: The circuit output is measured to produce the full computational-basis probability distribution ( $2^4 = 16$ probabilities) rather than a compressed expectation value.
Experimental Controls:
- Readout Ladder: Compared Full-Probs against variants including Expval-Z (4 features), Z+ZZ (pairwise correlations), and truncated probability vectors (Prob-k4, Prob-k8).
- Baseline Models: Classical softmax attention, classical MLP-value attention, ResNet18, and ViT-Tiny.
- Frozen-Random Control: A configuration where the quantum circuit parameters are randomly initialized and frozen, while classical layers remain trainable, to test if the benefit stems from learned quantum dynamics or the representation capacity itself.
- Diagnostic Tests: Included probability-vector compression (PCA), frozen readout-only probes, and a mixed-state depolarizing channel stress test.
Protocol: Experiments were run with ten random seeds. Statistical significance was assessed using paired Wilcoxon signed-rank tests with Holm correction, and practical equivalence was evaluated using Two One-Sided Tests (TOST) with a 0.02 AUC margin.

Key Contributions

Architectural Proposal: Introduction of a hybrid quantum self-attention mechanism where the value path utilizes full computational-basis probability readout, contrasting with the standard expectation-value approach.
Controlled Readout Analysis: A systematic "readout ladder" experiment isolating the measurement map as the primary variable, demonstrating that different measurement strategies on the same circuit yield different performance outcomes.
Rigorous Statistical Evaluation: A multi-seed evaluation across three datasets employing paired statistical tests, equivalence testing, and diagnostic probes to distinguish between representation capacity and quantum optimization effects.
Methodological Insight: Evidence suggesting that in small-qubit simulations, the choice of readout (measurement map) is a critical architectural factor that can materially alter model performance, independent of the quantum circuit's training dynamics.

Results

Performance Gains: TE-HQViT-Probs outperformed its expectation-value counterpart (TE-HQViT-Expval) in mean Area Under the Curve (AUC) across all three datasets.
- PneumoniaMNIST: +0.0062 AUC (small, nominal significance).
- BreastMNIST: +0.0497 AUC (large effect, near corrected threshold).
- OrganAMNIST: +0.0236 AUC (large, statistically significant after correction).
Readout Ladder: The Full-Probs configuration achieved the highest mean AUC among all readout variants (Expval-Z, Z+ZZ, Prob-k4, Prob-k8) on all datasets, confirming that the full probability vector preserves information lost by compressed observables.
Classical Comparison: While TE-HQViT-Probs improved over its quantum expectation-value variant, it did not universally surpass classical baselines. ResNet18 achieved the highest AUC on all datasets, and TE-HQViT-Probs was inferior to classical softmax attention on BreastMNIST.
Diagnostic Findings:
- Frozen-Random Control: Trainable and frozen-random probability readouts performed similarly, suggesting the improvement is driven by the nonlinear feature map of the amplitude embedding and probability readout rather than the optimization of quantum parameters.
- Compression: A 4-dimensional PCA projection of the 16-dimensional probability vector outperformed the native 4-dimensional Expval-Z vector, indicating the benefit is not merely due to higher dimensionality but the specific information content of the probabilities.
- Noise: A depolarizing channel stress test showed a slight AUC decrease under simulated noise, indicating limited robustness in this specific simulation context.

Significance and Claims
The paper concludes with a methodological rather than computational significance. The primary finding is that full-probability readout can preserve class-relevant information discarded by simple expectation-value readout in small-qubit hybrid models. The author explicitly states that:

The results do not establish quantum computational advantage, hardware speedup, or readiness for clinical deployment.
The observed performance gains are attributed to representation and measurement capacity rather than learned quantum dynamics alone.
Readout design should be treated as a core architectural variable in hybrid quantum attention research, comparable in importance to circuit depth or encoding strategies.
The study serves as a diagnostic upper-bound for small circuits, suggesting that future scalable approaches should investigate learned compressed probability readouts or adaptive observable sets to avoid the exponential scaling of full-probability readout ( $2^n$ ).

Technical Summary: TE-HQViT-Probs

More like this