A research team from India has demonstrated that integrating parameterized quantum circuits with classical convolutional neural networks can improve the computational detection of pulmonary tuberculosis on chest radiographs, according to a peer-reviewed evaluation published in Frontiers in Artificial Intelligence. Early clinical identification of pulmonary tuberculosis remains notoriously difficult because early parenchymal lesions often present as subtle, low-contrast structural anomalies that blend into surrounding anatomical tissue. Although conventional screening tools like sputum smear microscopy are accessible and widely deployed across endemic territories, their sensitivity varies widely between thirty and seventy percent while failing to distinguish drug-resistant strains. Chest radiography offers higher sensitivity, typically ranging between eighty-seven and ninety-eight percent according to World Health Organization reports, but radiographic presentations frequently overlap with other infectious and inflammatory conditions, resulting in compromised diagnostic specificity. To overcome these diagnostic hurdles, the investigators designed a hybrid computational framework coupling four-qubit quantum circuits with standard convolutional vision backbones, evaluating whether non-classical feature representations could refine the classification of chest X-ray images.

The study evaluated diagnostic efficacy using the public Shenzhen Hospital chest X-ray dataset curated by the United States National Library of Medicine, containing six hundred and sixty-two postero-anterior radiographs from Shenzhen No. 3 People’s Hospital in China. The cohort exhibits an evenly distributed demographic representation comprising three hundred and twenty-six normal control cases and three hundred and thirty-six confirmed tuberculosis-positive radiographs, yielding a baseline class ratio of approximately zero point ninety-seven to one that mitigates algorithmic bias stemming from class imbalance. Classical deep learning classifiers have attained expert-level diagnostic precision in medical imaging tasks, yet standard convolutional neural networks often exhibit brittle generalization when exposed to limited, heterogeneous, or noisy training cohorts, such as pediatric presentations or cases co-occurring with human immunodeficiency virus. Quantum machine learning introduces alternative computational paradigms by mapping classical inputs into multi-qubit Hilbert spaces, where principles of quantum superposition and quantum entanglement allow parameterized architectures to capture non-linear structural correlations that may remain inaccessible to standard purely classical convolutional filters.

To construct the hybrid processing pipeline, the researchers combined parameterized quantum circuits with six established classical convolutional neural network backbones spanning diverse topological designs, namely VGG16, ResNet50, DenseNet121, InceptionV3, EfficientNetB0, and EfficientNetV2-S. Each classical architecture was initialized using transfer learning weights pre-trained on the ImageNet database, capitalizing on general visual representations optimized across millions of natural images. In this dual-branch design, the classical backbone extracts high-dimensional spatial embeddings from full-resolution chest radiographs resized to two hundred and fifty-six by two hundred and fifty-six pixels, isolating deep semantic patterns from the penultimate layer of the neural network. Concurrently, an independent classical preprocessing pathway prepares a coarse projection for quantum circuit ingestion, downsampling the enhanced grayscale radiograph into a compact two-by-two spatial pixel matrix. This aggressive reduction produces four scalar values designed to capture broad global structural contrast rather than fine textural microstructures, ensuring dimensional compatibility with near-term quantum processing capabilities.

The mathematical transformation within the quantum branch was executed using the PennyLane quantum simulation framework configured for a four-qubit register. The four preprocessed image scalar values were z-score normalized and bounded within the interval from minus pi to positive pi to conform to the angular rotational domain of quantum operators. The framework then utilized angle embedding, mapping each scalar feature directly to the rotation angle of a single-qubit Pauli-Y logic gate applied across the four respective qubits. This operation transfers classical visual information into the probabilistic amplitudes of a multi-qubit quantum state. Immediately following the initial angle encoding layer, the circuit deploys an entangling unitary sequence composed of three sequential Controlled-NOT gates organized in a linear chain topology connecting qubit zero to one, one to two, and two to three. This linear entanglement creates inter-qubit state correlations, ensuring that the feature information encoded into each individual qubit propagates downstream through the quantum state vector without demanding exhaustive all-to-all physical hardware connectivity.

Feature extraction from the evolved quantum state was achieved through observable measurements inspired by Hamiltonian mechanics. Rather than simulating physical Hamiltonian time evolution during the data encoding stage, the researchers applied a weighted multi-qubit Hamiltonian measurement operator defined by coefficients across Pauli operators, specifically zero point two on Pauli-X for qubit zero, minus zero point fifty-four on Pauli-Y for qubit one, plus zero point twelve on Pauli-Z for qubit two, and minus zero point twenty on Pauli-X for qubit three. The circuit measures the expectation values of these Pauli operators, projecting the complex high-dimensional quantum Hilbert space onto fixed-dimensional real values. In parallel formulations, the single-qubit expectation values of Pauli-Z operators were extracted to form a concise quantum feature vector. This generated quantum vector was subsequently concatenated with the deep high-dimensional latent vector generated by the classical convolutional backbone. The combined hybrid vector was then routed through a classical multilayer perceptron activated by rectified linear units and culminated in a single sigmoid-activated output neuron predicting the probability of pulmonary tuberculosis.

Rigorous experimental validation was implemented across all model families using five-fold stratified cross-validation, preserving the balanced ratio of healthy and pathological cases across all testing splits. Training routines were conducted on central processing unit workstations over fifty epochs per fold using the AdamW optimization algorithm, with a batch size of sixteen, an initial learning rate of one times ten to the negative three, a ten percent linear warmup schedule, and cosine annealing learning rate decay. For quantum hybrid models, the base learning rate was multiplied by zero point three to stabilize gradient propagation across the quantum circuit layers, while gradient clipping with a maximum norm threshold of one point zero was enforced to prevent numerical explosion during parameter backpropagation. Binary cross-entropy with logits served as the training loss function across all configurations, ensuring uniform optimization criteria across classical baselines, quantum-augmented hybrids, and controlled ablation models.

To isolate whether observed empirical improvements originated from genuine quantum processing rather than mere structural parameter expansion, the authors executed a controlled ablation analysis. In each ablated configuration, the parameterized quantum circuit was entirely removed and substituted with a classical Sobel-texture feature extraction pathway coupled to a shallow multilayer perceptron. This ablated baseline retained the identical classical convolutional neural network backbone, identical downstream multilayer perceptron fusion head, and identical intermediate feature dimensionality, replacing only the quantum state preparation, entangling Controlled-NOT gates, and quantum expectation measurements with classical edge-gradient texture transformations. All ablated architectures were trained using the identical cross-validation folds, dataset splits, optimization parameters, and learning schedules. This meticulous architectural control allowed researchers to isolate whether the hybrid performance gains represented true non-classical mathematical synergies or could be replicated by conventional image processing routines.

Under five-fold cross-validation, this quantum-hybrid model attained an area under the receiver operating characteristic curve of zero point nine six eight six plus or minus zero point zero zero seven four, accompanied by a ninety-five percent confidence interval ranging from zero point nine three four to zero point nine six eight. Its diagnostic sensitivity reached zero point nine three one five, paired with a specificity of zero point eight six two four, an overall accuracy of zero point eight nine seven three, and an F1-score of zero point nine zero two seven. By comparison, the classical standalone EfficientNetB0 classifier achieved an area under the curve of zero point nine six six eight, an accuracy of zero point nine zero three two, a sensitivity of zero point nine one three five, and a specificity of zero point eight nine two seven, while the ablated counterpart achieved an area under the curve of zero point nine six five eight.

Pairwise statistical evaluations using McNemar tests on paired cross-validation predictions revealed nuanced insights into error distribution shifts across model paradigms. The quantum-hybrid EfficientNetB0 exhibited statistically significant classification differences when contrasted against several classical baselines, including classical ResNet50 with a chi-square of four point five zero and p-value of zero point zero three four, classical DenseNet121 with a chi-square of one hundred and fifty-seven point nine two four and p-value below zero point zero zero one, and classical VGG16 with a chi-square of two hundred and ninety-six point six nine zero and p-value below zero point zero zero one. Furthermore, quantum-hybrid EfficientNetB0 differed significantly from the classical EfficientNetB0 baseline with a chi-square of eighteen point eight nine three and p-value below zero point zero zero one. However, the direct McNemar comparison between the quantum-hybrid EfficientNetB0 and its classical Sobel-ablated counterpart yielded a chi-square of zero point zero seven one with a p-value of zero point seven eight nine. Because this difference failed to achieve statistical significance at the nominal threshold of zero point zero five, the authors noted that the observed performance boost represents an incremental enhancement rather than definitive proof of a transformative quantum advantage.

Computational training duration exhibited substantial variance across distinct convolutional backbones when coupled to simulated quantum circuits. Compact architectures maintained manageable overhead, with the quantum-hybrid EfficientNetB0 requiring sixteen thousand eight hundred and eighty-five seconds, equivalent to roughly four point seven hours, which proved faster than both its ablated counterpart at seventeen thousand seven hundred and fifty-three seconds and its classical baseline at twenty-three thousand nine hundred and forty-six seconds. Conversely, heavier backbones suffered pronounced computational penalties. The quantum-hybrid VGG16 required one hundred and one thousand two hundred and seven seconds, or approximately twenty-eight point one hours, more than doubling the training time of both its classical and ablated versions due to expansive intermediate activation tensors interacting unfavorably with PennyLane’s parameter-shift differentiation rule. Similarly, the multi-branch structure of InceptionV3 expanded gradient computation graphs, requiring thirty thousand nine hundred and fifty seconds. The researchers emphasized key study limitations, including reliance on simulated CPU quantum environments rather than physical noisy intermediate-scale quantum hardware, information loss induced by downsampling radiographs to four-pixel inputs, and restriction to binary classification without multi-disease stratification.

Subject of Research: Hybrid quantum-classical deep learning architectures for pulmonary tuberculosis detection on chest radiographs

Article Title: Hamiltonian-encoded quantum feature maps augment classical CNNs for pulmonary tuberculosis detection

Article References: Thomas, S. D., Rajan, A., Varghese, L., Joy, N., Balakrishnan, A., Subair, S., Shajan, I., Shahul, M., Vinodchandra, S. S., Abhithaj, J., Devdas Acharya, K., Vijayakumar, M., Kandel Codi, J. A., & Raju, R. (2026). Hamiltonian-encoded quantum feature maps augment classical CNNs for pulmonary tuberculosis detection. Frontiers in Artificial Intelligence, 9, Article 1887124. https://doi.org/10.3389/frai.2026.1887124

Image Credits: AI Generated

DOI: 10.3389/frai.2026.1887124

Keywords: quantum machine learning, pulmonary tuberculosis, chest radiography, hybrid neural networks, convolutional neural networks, EfficientNetB0, parameterized quantum circuits, angle embedding, entangling gates, Hamiltonian feature extraction, Sobel ablation study, McNemar test, Shenzhen Hospital dataset, computer-aided detection, medical imaging, PennyLane simulation