Quantum machine learning achieves 86.4% accuracy in detecting leukemia in 50 samples

Detecting acute myeloid leukemia (AML) from microscopic images of blood cells is a major challenge, but new research shows quantum machine learning may provide a viable solution. A. Bano and L. Liebovitch, together with colleagues, demonstrate the potential of algorithms such as balanced propagation and variational quantum circuits to accurately identify AML, even with limited data and computational resources. Their work utilizing the AML-cell morphology dataset achieved surprisingly competitive results, reaching up to 86.4% accuracy with equilibrium propagation, despite lower image resolution and fewer training samples than traditional methods. This work establishes an important baseline for near-term quantum computing applications in healthcare and suggests that quantum machine learning remains a viable approach for medical image analysis even in the NISQ era.

This study utilized the AML Cell Morphology Dataset, a collection of 18,365 expert-annotated images, focused on establishing a reproducible baseline for QML in medicine, and validated the potential of QML in the current noisy intermediate-scale quantum (NISQ) era. The team achieved this by training the model on a limited subset of the dataset, ranging from 50 to 250 samples per class, and intentionally introducing constraints such as reduced image resolution (64×64 pixels) and engineering features (20D) to simulate realistic computational scenarios.

Experiments have shown that these quantum methods achieved performance levels that are only 12-15% lower than classical convolutional neural networks (CNNs), a surprising result considering the limitations imposed. Specifically, the EP algorithm, whose state collapse measurements decisively avoid backpropagation, a process incompatible with quantum systems, reached an accuracy of 86.4%, just 12% below CNN’s performance. Four-qubit VQC utilizing ZZFeatureMap encoding and shallow RealAmplitudes analysis achieved 83.0% accuracy and demonstrated consistent data efficiency. CNN required 5 times more data, 250 samples, to achieve 98% accuracy, whereas CNN maintained a stable 83% performance with just 50 samples per class. This data efficiency is particularly noteworthy since expert annotations in the medical field are often expensive and time-consuming to obtain.
In this study, we leverage artificial image features and training without backpropagation to establish a reproducible EP pipeline for AML detection and present benchmarks across various dataset sizes to evaluate QML performance and execution time. By running these simulations on a standard laptop computer using the IBM Qiskit quantum simulator, the scientists demonstrated that even exploratory studies can gain valuable insight into the potential of these new QML algorithms. This work addresses key challenges in the field, including scaling beyond toy datasets, handling real-world data fluctuations, and demonstrating practical advantages over classical methods, paving the way for future advances in quantum-enhanced medical diagnostics. To facilitate this, we leveraged the AML Cell Morphology Dataset, a collection of 18,365 expert-annotated images and strategically selected subsets ranging from 50 to 250 samples per class for training and testing. In the experiments, a rigorous methodology including image preprocessing and feature engineering was adopted to reduce the computational complexity, images were downscaled to 64 × 64 pixels, and a 20-dimensional feature space was designed from the original data.

The team then implemented EP, an energy-based learning method that uses state collapse measurements to avoid the need for backpropagation, a process that is incompatible with quantum systems. EP was trained to classify AML cells without relying on gradients, but instead utilizing equilibrium states derived from free and nudge dynamics. This is an important innovation for potential quantum hardware implementations. Additionally, this work pioneered a four-qubit VQC classifier that leverages ZZFeatureMap for encoding classical data into quantum states and shallow RealAmplitudes analysis for circuit parameter optimization. This VQC was designed to maintain consistent performance even with limited data and demonstrated its data efficiency, achieving 83% accuracy with just 50 samples per class. On the other hand, traditional CNN requires 250 samples to reach 98% accuracy.

All quantum simulations were performed using the IBM Qiskit platform on a standard laptop computer to establish a reproducible baseline for QML in healthcare. The results revealed that EP achieved 86.4% accuracy, only 12% below the performance of traditional CNNs, while VQC achieved 83.0% accuracy, showing competitive performance despite operating under severe constraints. This study examines the potential of QML for short-term applications in healthcare and provides a path to leveraging quantum resources to improve medical diagnostics.

Detecting AML with quantum machine learning

Scientists have achieved competitive performance using quantum machine learning (QML) algorithms in real-world medical image processing, even under severe constraints. Key to this study was to rigorously test the feasibility of QML in a medical setting using an AML cell morphology dataset containing 18,365 images annotated by experts. Experiments reveal that the quantum method can achieve 12-15% lower performance levels than classical convolutional neural networks (CNNs), despite operating at a reduced image resolution (64×64 pixels) and exploiting engineered features limited to 20 dimensions.

Importantly, the EP algorithm reached 86.4% accuracy without using backpropagation. Backpropagation is a technique that is incompatible with quantum systems due to state collapse measurements and shows a 12% performance gap compared to CNN benchmarks. The data show that the 4-qubit VQC achieves 83.0% accuracy and exhibits consistent data efficiency. We maintained a stable 83% performance level using only 50 samples per class. Testing proves that VQC requires five times less data than CNN, which requires 250 samples to achieve 98% accuracy, highlighting its potential benefits in scenarios where labeled data is scarce. The team measured performance across a variety of dataset scales ranging from 50 to 250 samples per class to establish a reproducible baseline for QML in healthcare and validate its feasibility in the current Noisy Intermediate Scale Quantum (NISQ) era.

This breakthrough provides a functional EP pipeline for AML detection trained without backpropagation and a 4-qubit VQC classifier employing ZZFeatureMap encoding and shallow RealAmplitudes analysis. Measurements confirm that this study systematically compares quantum-inspired (EP) and pure quantum (VQC) approaches under identical experimental conditions, quantifying data efficiency, a critical factor for medical applications where expert annotation is costly and time-consuming. This study was performed using a laptop computer and the IBM Qiskit quantum simulator and demonstrated that exploratory simulations also provide valuable insight into the performance of these new QML algorithms. These results establish the foundation for future research exploring more complex QML architectures and their applications to a wide range of medical imaging challenges.

Detecting AML with quantum machine learning

Scientists have demonstrated the feasibility of applying quantum machine learning (QML) algorithms to real-world medical image processing tasks. Key results show that the quantum method achieved performance within 12-15% of classical convolutional neural networks (CNNs), despite operating with limited data (50-250 samples per class) and low image resolution. . In particular, EP reached 86.4% accuracy without utilizing backpropagation, a technique that is incompatible with quantum systems, while 4-qubit VQC achieved 83.0% accuracy and maintained consistent performance even with only 50 samples per class. This exceeds the data efficiency of CNN, which requires 250 samples to achieve peak performance.

This suggests a potential advantage for QML in scenarios where labeled data is scarce, such as rare disease research. The employed shallow circuit design is compatible with current Noisy Intermediate-Scale Quantum (NISQ) hardware and facilitates short-term verification. . The authors acknowledge limitations, including the current superiority of classical CNNs in absolute accuracy for this particular classification task. Future research will focus on shot-based sampling experiments, error mitigation strategies such as zero-noise extrapolation, analysis of barren plateaus during training, and the development of hybrid quantum-classical architectures. Further validation across diverse patient populations is planned. These findings establish a reproducible baseline for QML in healthcare and demonstrate that quantum methods may be a viable alternative when data availability rather than model capacity is the primary constraint.