Skip to main content
NCBI home page
NIHPA Author Manuscripts logoLink to NIHPA Author Manuscripts
. Author manuscript; available in PMC: 2026 Apr 27.
Published in final edited form as: Proc SPIE Int Soc Opt Eng. 2026 Apr 2;13930:1393019. doi: 10.1117/12.3085665

Prostate Cancer Classification Using Quantum Machine Learning on Multi-parametric MRI

Peng Chen a,b, Mojtaba Safari c, Rowan Barker-Clarke a, Xiaofeng Yang c,‡, Jacob G Scott a,b,d,e,†
PMCID: PMC13110830  NIHMSID: NIHMS2166829  PMID: 42046760

Abstract

Prostate cancer is one of the most common malignancies in men, and accurate classification of lesions into clinically significant or insignificant categories is essential for patient management. Multiparametric MRI, including T2-weighted (T2W) and apparent diffusion coefficient (ADC) imaging, enables extraction of quantitative radiomic features that can be exploited by machine learning for improved diagnosis. While classical machine-learning models such as support vector machines (SVM), random forests (RF), and extreme gradient boosting (XGBoost) have shown strong performance in radiomics-based classification, quantum machine learning offers a new paradigm that leverages quantum feature spaces, potentially uncovering complex patterns inaccessible to classical kernels. In this study, we systematically compared three classical classifiers (SVM-RBF, RF, and XGBoost) with three quantum support vector machine (QSVM) variants: amplitude encoding, angle encoding, and angle encoding with a projected quantum kernel, for classifying 299 prostate lesions from the PROSTATEx Challenge dataset. Radiomics features were extracted from T2W and ADC images. A nested stratified cross-validation pipeline was employed, with feature selection performed in each outer fold and hyperparameters optimized via grid search. QSVM-amplitude encoding achieved the highest mean AUC (0.799 ± 0.082), outperforming SVM-RBF (0.608 ± 0.244) and matching or exceeding RF (0.728 ± 0.083) and XGBoost (0.720 ± 0.065), while offering higher sensitivity at comparable specificity. These findings demonstrate that qubit-efficient QSVMs can deliver competitive or superior performance in small-sample, low-dimensional clinical imaging settings, highlighting their potential for prostate cancer lesion classification.

Keywords: Quantum machine learning, Multiparametric MRI, Radiomics, Prostate Cancer

1. INTRODUCTION

Prostate cancer is one of the most frequently diagnosed malignancies in men and a leading cause of cancer-related mortality worldwide.1, 2 Accurate classification of prostate lesions into clinically significant and non-clinically significant categories is critical for guiding patient management, limiting unnecessary biopsies, and optimizing treatment strategies.3, 4 Multiparametric magnetic resonance imaging (mpMRI), incorporating anatomical and functional sequences such as T2-weighted (T2W) imaging and apparent diffusion coefficient (ADC) maps, has emerged as a powerful non-invasive tool for detecting and characterizing prostate cancer.5–7 Beyond qualitative interpretation, quantitative radiomics enables the extraction of high-dimensional imaging features that capture subtle tissue heterogeneity and morphology, offering the potential to improve diagnostic performance through machine learning approaches.8–10

Classical machine learning (ML) methods, including support vector machines, ensemble tree models, and gradient boosting algorithms, have been widely applied to radiomics-based prostate cancer classification.11–14 These algorithms leverage engineered features to learn complex decision boundaries, often achieving strong performance when combined with appropriate feature selection and cross-validation frameworks. This approach aligns with successful applications in other oncology domains, such as predicting genetic mutations from MRI15 or assessing early treatment response using advanced imaging heterogeneity analysis.16 However, classical models are ultimately constrained by the representational capacity of their chosen kernels or ensemble structures, potentially limiting their ability to fully capture intricate nonlinear relationships in high-dimensional feature spaces.17

Quantum machine learning (QML) offers a novel paradigm by exploiting quantum mechanical principles such as superposition and entanglement to process information in a quantum feature space that may be exponentially larger than its classical counterpart.18 Quantum kernel methods, in particular, enable support vector machines to operate on similarity measures derived from quantum states prepared by parameterized circuits, potentially revealing patterns inaccessible to classical kernels.17, 19, 20 This paradigm is part of a broader trend towards increasingly sophisticated AI models in medicine, including the development of foundational vision-language models for multimodal biomedical data.21 Recent advances in quantum algorithms and simulators have spurred interest in exploring QML for medical imaging tasks, where complex, high-dimensional data structures are common.22–26

In this study, we investigate QML approaches for classifying prostate lesions using radiomics features extracted from mpMRI. We conduct a comparative analysis between classical and quantum ML models. Specifically, we evaluate three classical classifiers, including support vector machine, random forest, and extreme gradient boosting, and three quantum support vector machine variants with distinct feature embedding strategies (Figure 1(A)). Our goal is to assess whether quantum kernel methods can offer tangible performance advantages over established classical algorithms in this clinically meaningful application.

Figure 1.

Figure 1.

Open in a new tab

(A) Workflow for classical and quantum ML analysis of mpMRI radiomics for prostate lesion classification, with (B) representative quantum circuits for QSVM-amplitude, QSVM-angle, and QSVM-angle+PQK.

2. METHODS

2.1. Data acquisition

The mpMRI images used in this study were obtained from the PROSTATEx Challenge dataset, comprising 299 prostate lesions with both T2w images and ADC maps derived from diffusion-weighted imaging.27 All scans were acquired at the Prostate MR Reference Center, Radboud University Medical Centre, with acquisition details described in Litjens et al.28 Expert-annotated lesion contours for both T2w and ADC images were provided as binary masks by the Department of Advanced Biomedical Sciences, University of Naples “Federico II,” as previously reported in Cuocolo et al.29 Our classification of lesions follows the definitions provided in Donisi et al.,13 where Gleason grade group ≥2 was considered clinically significant (n=76), and Gleason grade group < 2 was considered non-clinically significant (n=223).

2.2. Radiomic Feature Extraction

Radiomic features were extracted using the open-source PyRadiomics library.30 As implemented in PyRadiomics, all images were resampled to an isotropic voxel spacing of 1 × 1 × 1 mm3 and intensity-normalized by centering voxel values to a mean of zero and scaling by a factor of 100. To standardize texture feature computation, gray levels were discretized using a fixed bin count of 32. Feature extraction was performed across all supported image types (original, wavelet, Laplacian of Gaussian, etc.) and all PyRadiomics feature classes, including first-order statistics, shape, gray-level co-occurrence matrix (GLCM), gray-level run length matrix (GLRLM), gray-level size zone matrix (GLSZM), neighboring gray tone difference matrix (NGTDM), and gray-level dependence matrix (GLDM). This process yields 1,706 quantitative features per imaging modality (T2W and ADC) for each lesion.

2.3. Feature Selection

Feature selection was performed within a nested stratified cross-validation (CV) framework to ensure unbiased performance estimation and to prevent information leakage. The outer loop consisted of five folds, with each fold serving once as an independent test set. Within each outer training split, a Balanced Random Forest classifier was trained using an inner CV loop (four folds) to compute feature importances. For each feature, the mean importance score was calculated across the inner folds, and the eight highest-ranking features were selected for model training. The process was repeated independently for each outer fold.

2.4. Classical Machine Learning Models

Three widely used classical classifiers were evaluated: Support Vector Machine with Radial Basis Function kernel (SVM-RBF), Random Forest (RF), and Extreme Gradient Boosting (XGBoost). The SVM-RBF is a kernel-based model that maps input features into a high-dimensional space using a Gaussian kernel and finds the optimal separating hyperplane; the penalty parameter C was tuned over {0.01, 0.1, 1, 10, 100} and the kernel width parameter γ over {10−4, 10−3, 10−2, 10−1, 1} The RF is an ensemble of decision trees trained on bootstrap samples with random feature selection at each split to reduce overfitting; grid search covered the number of trees {300, 500}, maximum depth {None, 8, 12}, and minimum samples per leaf {1, 2, 4}. XGBoost is a gradient boosting algorithm that builds trees sequentially, with each new tree correcting the residual errors of the ensemble; grid search included the number of trees {300, 500}, maximum depth {3, 5}, learning rate {0.05, 0.1}, subsample fraction {0.8, 1.0}, column subsample fraction per tree {0.8, 1.0}, minimum child weight {1, 3}, and L2 regularization parameter λ ∈ {1, 3}. All classical models were implemented within a nested stratified CV framework, with the outer five folds used for independent performance evaluation and the inner folds for exhaustive grid search hyperparameter tuning; before training, selected features in each outer fold were standardized to zero mean and unit variance based on the training set statistics.

2.5. Quantum Machine Learning Models

Quantum machine learning models were implemented in PennyLane using the same selected features and nested cross-validation design as for the classical models.31 All quantum classifiers in this study were Quantum Support Vector Machines (QSVMs), which are analogous to classical SVMs but replace the standard kernel function with a quantum kernel computed from the inner product of quantum states prepared by a parameterized quantum circuit. This approach allows the SVM to operate in a quantum feature space that may be exponentially large, potentially capturing complex data relationships not easily accessible to classical kernels.

QSVM-amplitude encodes the entire feature vector into the amplitude of a quantum state (amplitude embedding). The input features were first normalized to unit length (row-wise L2 normalization) before encoding. This embedding is qubit-efficient, requiring only log2(nfeat.) qubits, which for 8 features equals 3 qubits. All features are encoded simultaneously in a single layer without explicit rotation gates.

QSVM-angle uses an angle embedding scheme in which each feature is encoded as a rotation angle applied to a dedicated qubit. The number of qubits equals the number of selected features per fold (8 qubits). The encoding-and-entanglement block can be repeated for greater expressivity (feature map depth ∈ {1, 2, 3}), and a rotation scale parameter α ∈ {0.5, 1, 2} controls the magnitude of the feature-dependent rotations.

QSVM-angle+PQK extends the QSVM-angle model by adding projected quantum kernel (PQK) entangling layers, which apply controlled rotation gates between adjacent qubits with a tunable entanglement strength ∈ {0.25, 0.5, 1.0}. This extra entanglement is intended to capture higher-order correlations between features.

For all QSVM variants, the similarity between two samples was computed as the squared magnitude of the inner product between their respective quantum states, and the resulting quantum kernel matrix was used to train the SVM. The penalty parameter C was tuned over {0.01, 0.1, 1, 10, 100} using the inner folds of the nested CV framework. All quantum circuits (Figure 1(B)) were executed using the default.qubit statevector simulator backend in PennyLane, providing idealized, noise-free simulators of quantum hardware.

2.6. Evaluation Criteria

Model performance was evaluated over the five outer cross-validation folds using the area under the receiver operating characteristic curve (AUC), along with accuracy, sensitivity, specificity, and F1 score. The mean and standard deviation (mean ± SD) of all metrics, including AUC, were reported in the results table to summarize performance and variability across folds. In addition, all out-of-fold decision scores were combined to generate a pooled receiver operating characteristic curve and corresponding pooled AUC for each model.

3. RESULTS

3.1. Feature Selection Stability

The nested cross-validation feature selection process demonstrated high stability, consistently identifying radiomic features derived from the ADC map. As detailed in Table 1, five features were selected across all five outer folds. The most stable features were primarily first-order statistics and texture features from the GLDM and GLRLM matrices, predominantly extracted from mathematically transformed ADC images (square, logarithm). This consistency underscores the importance of ADC-derived measures of tissue heterogeneity and intensity distribution for distinguishing clinically significant prostate lesions.

Table 1.

Stability of selected radiomic features across outer folds.

Feature Selected folds
ADC_square_gldm_LargeDependenceLowGrayLevelEmphasis 5/5
ADC_square_glrlm_LongRunLowGrayLevelEmphasis 5/5
ADC_square_gldm_LowGrayLevelEmphasis 5/5
ADC_logarithm_firstorder_InterquartileRange 5/5
ADC_logarithm_firstorder_RobustMeanAbsoluteDeviation 5/5
ADC_wavelet-LLL_firstorder_10Percentile 4/5
ADC_squareroot_firstorder_InterquartileRange 4/5
ADC_square_glrlm_LowGrayLevelRunEmphasis 3/5
ADC_squareroot_firstorder_RobustMeanAbsoluteDeviation 2/5
ADC_wavelet-LLL_firstorder_Median 1/5
ADC_square_firstorder_Median 1/5
Open in a new tab

3.2. Comparative Performance of Classical and Quantum Models

The performance of all evaluated classifiers is summarized in Table 2 and visualized in the pooled ROC curves (Figure 2). Compared to the classical SVM-RBF, all QSVM variants demonstrated markedly higher discrimination. The QSVM employing amplitude encoding achieved the highest mean AUC of 0.799 ± 0.082, substantially outperforming the classical SVM-RBF, which yielded an AUC of 0.608 ± 0.244. This superior performance of the quantum kernel methods was consistent, with QSVM-angle+PQK and QSVM-angle also attaining strong AUCs of 0.764 ± 0.075 and 0.754 ± 0.081, respectively.

Table 2.

Classification results of classical and quantum ML classifiers on selected 8 features.

Model AUC Accuracy Sensitivity Specificity F1
SVM 0.608 ± 0.244 0.767 ± 0.041 0.435 ± 0.249 0.882 ± 0.070 0.437 ± 0.235
Random Forest 0.728 ± 0.083 0.774 ± 0.051 0.488 ± 0.157 0.873 ± 0.042 0.517 ± 0.125
XGBoost 0.720 ± 0.065 0.777 ± 0.007 0.408 ± 0.045 0.905 ± 0.022 0.483 ± 0.028
QSVM-amplitude 0.799 ± 0.082 0.770 ± 0.040 0.528 ± 0.155 0.855 ± 0.047 0.531 ± 0.116
QSVM-angle 0.754 ± 0.081 0.777 ± 0.044 0.607 ± 0.074 0.836 ± 0.039 0.584 ± 0.077
QSVM-angle+PQK 0.764 ± 0.075 0.770 ± 0.023 0.567 ± 0.042 0.841 ± 0.032 0.559 ± 0.033
Open in a new tab

Figure 2.

Figure 2.

Open in a new tab

Pooled ROC curves from out-of-fold predictions comparing classical and quantum classifiers.

When compared against the more robust classical ensemble methods, the QSVMs remained highly competitive. The leading QSVM-amplitude model matched or slightly exceeded the discrimination achieved by Random Forest (AUC: 0.728 ± 0.083) and Extreme Gradient Boosting (XGBoost, AUC: 0.720 ± 0.065). Beyond aggregate discrimination, the QSVM models, particularly QSVM-angle, offered a favorable balance of metrics by delivering higher sensitivity. QSVM-angle achieved a mean sensitivity of 0.607 ± 0.074, compared to 0.488 ± 0.157 for Random Forest and 0.408 ± 0.045 for XGBoost, while maintaining comparable specificity (QSVM-angle: 0.836 ± 0.039; Random Forest: 0.873 ± 0.042; XGBoost: 0.905 ± 0.022). Overall accuracy was consistent across all models, clustering near 0.77.

4. DISCUSSION

This study presents a comparative analysis of classical and quantum machine learning models for classifying clinically significant prostate cancer using radiomic features from multiparametric MRI. The key finding is that QSVMs, particularly the qubit-efficient amplitude-encoding variant, achieved highly competitive performance, matching or exceeding that of established classical models like Random Forest and XGBoost. The superior performance of QSVM-amplitude over the classical SVM-RBF is especially notable, suggesting that the quantum feature space induced by the amplitude-encoding circuit can capture more discriminative patterns for this task than a classical radial basis function kernel. This provides a practical demonstration that quantum kernel methods can offer tangible advantages in a real-world, low-dimensional clinical imaging problem.

The performance of the QSVM models, which operated on a compact set of eight engineered radiomic features, aligns with a broader trajectory in medical AI toward learning more powerful representations from imaging data. While our study utilized hand-crafted features, the field is increasingly moving toward foundation models capable of learning these representations directly from raw images. For instance, recently developed vision foundation models for 3D medical imaging, such as Triad,32 demonstrate the potential to extract generalized, transferable imaging features at scale. In this context, the quantum kernels evaluated here could be seen as a complementary, novel computational paradigm for learning complex relationships, not from raw pixels, but from high-level, semantically rich feature vectors—whether they are engineered radiomics or embeddings from a foundation model. Future work exploring the synergy between foundation model-derived embeddings and quantum kernel methods represents a promising frontier for quantum machine learning in medicine.

Our results also highlight important practical considerations. The strong performance of the amplitude-encoding QSVM, which requires only log2(nfeat) qubits, is highly encouraging for near-term quantum computing applications where qubit count is limited. This efficiency makes it a viable candidate for early experimentation on real quantum hardware. However, the comparable performance of sophisticated classical ensemble methods like XGBoost underscores that quantum advantage is not absolute and is highly context-dependent. The choice of feature encoding was critical; the angle-encoding schemes, while more intuitive, did not outperform the amplitude encoding in this setting. Furthermore, the high sensitivity of the QSVM-angle model, which came at only a modest cost to specificity, suggests a potential clinical utility where identifying most true positive cases is prioritized.

Several limitations of this pilot study must be acknowledged. First, the analysis was conducted on a single public dataset using idealized quantum simulators. Performance on real, noisy quantum processors may differ and requires dedicated investigation. Second, the dimensionality of the feature space was deliberately kept low due to computational constraints of classical simulation; the scalability of these quantum kernels to the high-dimensional feature sets common in deep learning remains an open question. Third, the clinical validation is preliminary, and the binary classification task, while standard, simplifies the spectrum of prostate cancer pathology.

5. CONCLUSIONS

This work provides evidence that quantum kernel methods, specifically QSVMs with efficient encoding strategies, can perform competitively for prostate cancer classification from MRI radiomics. These findings contribute to the growing body of research exploring quantum algorithms for medical imaging. Future directions should focus on external validation across diverse datasets, integration with state-of-the-art feature extractors like medical imaging foundation models, and rigorous benchmarking on emerging quantum hardware to assess real-world utility and scalability.

ACKNOWLEDGMENTS

This research is supported in part by the National Institutes of Health under award numbers U54CA274513.

REFERENCES

  • [1].Siegel RL, Kratzer TB, Giaquinto AN, Sung H, and Jemal A, “Cancer statistics, 2025,” CA: A Cancer Journal for Clinicians 75(1), 10–45 (2025). [DOI] [PMC free article] [PubMed] [Google Scholar]
  • [2].Fonteyne V, Tree A, Castro E, Touijer K, and Walz J, “Prostate cancer,” The Lancet (2025). Published online 2025; Seminar reviewing pathology, diagnostic strategies, and treatments for prostate cancer. [Google Scholar]
  • [3].Matoso A and Epstein JI, “Defining clinically significant prostate cancer on the basis of pathological findings,” Histopathology 74(1), 135–145 (2019). [DOI] [PubMed] [Google Scholar]
  • [4].Malik A, Srinivasan S, and Batra J, “A new era of prostate cancer precision medicine,” Frontiers in Oncology 9, 1263 (2019). [DOI] [PMC free article] [PubMed] [Google Scholar]
  • [5].Weinreb JC et al. , “Pi-rads prostate imaging–reporting and data system: 2015, version 2,” European Urology (2016). [Google Scholar]
  • [6].Turkbey B, Pinto PA, Mani H, Bernardo M, Pang Y, McKinney YL, Khurana K, Ravizzini GC, Albert PS, Merino MJ, and Choyke PL, “Prostate cancer: value of multiparametric mr imaging at 3 t for detection–histopathologic correlation,” Radiology 255, 89–99 (Apr. 2010). [DOI] [PMC free article] [PubMed] [Google Scholar]
  • [7].Luczynńska E, Heinze-Paluchowska S, Domalik A, Cwierz A, Kasperkiewicz H, Blecharz P, and Jereczek-Fossa B, “The utility of diffusion weighted imaging (dwi) using apparent diffusion coefficient (adc) values in discriminating between prostate cancer and normal tissue,” Polish Journal of Radiology 79, 450–455 (Dec. 2014). [DOI] [PMC free article] [PubMed] [Google Scholar]
  • [8].Lambin P et al. , “Radiomics: extracting more information from medical images using advanced feature analysis,” European Journal of Cancer (2012). [Google Scholar]
  • [9].Gillies RJ, Kinahan PE, and Hricak H, “Radiomics: Images are more than pictures, they are data,” Radiology 278, 563–577 (Feb. 2016). [DOI] [PMC free article] [PubMed] [Google Scholar]
  • [10].Fehr D, Veeraraghavan H, Wibmer A, Gondo T, Matsumoto K, Vargas HA, Sala E, Hricak H, and Deasy JO, “Automatic classification of prostate cancer gleason scores from multiparametric magnetic resonance images,” Proceedings of the National Academy of Sciences of the United States of America 112, E6265–E6273 (Nov. 2015). [DOI] [PMC free article] [PubMed] [Google Scholar]
  • [11].Chaddad A, Niazi T, Probst S, Bladou F, Anidjar M, and Bahoric B, “Predicting Gleason Score of Prostate Cancer Patients Using Radiomic Analysis,” Frontiers in Oncology 8, 630 (2018). [DOI] [PMC free article] [PubMed] [Google Scholar]
  • [12].Chen T, Li M, Gu Y, Zhang Y, Yang S, Wei C, Wu J, Li X, Zhao W, and Shen J, “Prostate Cancer Differentiation and Aggressiveness: Assessment With a Radiomic-Based Model vs. PI-RADS v2,” Journal of Magnetic Resonance Imaging 49(3), 875–884 (2019). [DOI] [PMC free article] [PubMed] [Google Scholar]
  • [13].Donisi L, Cesarelli M, Castaldo A, De Lucia DR, Nessuno F, Spadarella G, and Ricciardi C, “A Combined Radiomics and Machine Learning Approach to Distinguish Clinically Significant Prostate Lesions on a Publicly Available MRI Dataset,” Journal of Imaging 7(10), 215 (2021). [DOI] [PMC free article] [PubMed] [Google Scholar]
  • [14].Zheng Y, Zhang J, Huang D, Hao X, Qin W, and Liu Y, “Detecting mri-invisible prostate cancers using a weakly supervised deep learning model,” International Journal of Biomedical Imaging, 1–12 (2024). [Google Scholar]
  • [15].Safari M, Beiki M, Ameri A, Toudeshki SH, Fatemi A, and Archambault L, “Shuffle-resnet: Deep learning for predicting lgg idh1 mutation from multicenter anatomical mri sequences,” Biomedical Physics & Engineering Express 8(6), 065036 (2022). [Google Scholar]
  • [16].Beigi M, Kazerooni AF, Safari M, Alamolhoda M, Moghdam MS, Moghadam S, SalighehRad H, and Ameri A, “Heterogeneity analysis of diffusion-weighted mri for prediction and assessment of microstructural changes early after one cycle of induction chemotherapy in nasopharyngeal cancer patients,” La radiologia medica 123(1), 36–43 (2018). [DOI] [PubMed] [Google Scholar]
  • [17].Havlíček V, Córcoles AD, Temme K, Harrow AW, Kandala A, Chow JM, and Gambetta JM, “Supervised learning with quantum-enhanced feature spaces,” Nature 567, 209–212 (2019). [DOI] [PubMed] [Google Scholar]
  • [18].Biamonte J, Wittek P, Pancotti N, Rebentrost P, Wiebe N, and Lloyd S, “Quantum machine learning,” Nature 549(7671), 195–202 (2017). [DOI] [PubMed] [Google Scholar]
  • [19].Rebentrost P, Mohseni M, and Lloyd S, “Quantum Support Vector Machine for Big Data Classification,” Physical Review Letters 113(13), 130503 (2014). Published 25 September 2014; Received 12 February 2014. [DOI] [PubMed] [Google Scholar]
  • [20].Jäger J and Krems RV, “Universal expressiveness of variational quantum classifiers and quantum kernels for support vector machines,” Nature Communications 14, 576 (2023). [Google Scholar]
  • [21].Wang S, Jin Z, Hu M, Safari M, Zhao F, Chang C-W, Qiu RL, Roper J, Yu DS, and Yang X, “Unifying biomedical vision-language expertise: Towards a generalist foundation model via multi-clip knowledge distillation,” arXiv preprint arXiv:2506.22567 (2025). [Google Scholar]
  • [22].Gupta RS, Wood CE, Engstrom T, Pole JD, and Shrapnel S, “A systematic review of quantum machine learning for digital health,” npj Digital Medicine 8, 237 (2025). [DOI] [PMC free article] [PubMed] [Google Scholar]
  • [23].Felefly T, Roukoz C, Fares G, Achkar S, Yazbeck S, Meyer P, Kordahi M, Azoury F, Nehme D, Nasr E, Noël G, and Francis Z, “An explainable mri-radiomic quantum neural network to differentiate between large brain metastases and high-grade glioma using quantum annealing for feature selection,” Journal of Digital Imaging 36(6), 2335–2346 (2023). [DOI] [PMC free article] [PubMed] [Google Scholar]
  • [24].Moradi S, Spielvogel C, Krajnc D, Brandner C, Hillmich S, Wille R, Traub-Weidinger T, Li X, Hacker M, Drexler W, and Papp L, “Error mitigation enables pet radiomic cancer characterization on quantum computers,” European Journal of Nuclear Medicine and Molecular Imaging 50(13), 3826–3837 (2023). [DOI] [PMC free article] [PubMed] [Google Scholar]
  • [25].Buvaneswari T, Ramkumar M, Venkatesan P, and Sarath Kumar R, “Leveraging radiomics and hybrid quantum–classical convolutional networks for non-invasive detection of microsatellite instability in colorectal cancer,” Molecular Imaging and Biology 27(2), 227–237 (2025). [DOI] [PubMed] [Google Scholar]
  • [26].Huang X, Xu F, Zhu W, Yao L, He J, Su J, Zhao W, and Hu H, “An integrated strategy based on radiomics and quantum machine learning: Diagnosis and clinical interpretation of pulmonary ground-glass nodules,” BMC Medical Imaging 25(1), 279 (2025). [DOI] [PMC free article] [PubMed] [Google Scholar]
  • [27].Litjens G, Debats O, Barentsz J, Karssemeijer N, and Huisman H, “SPIE-AAPM PROSTATEx challenge data (version 2) [dataset].” The Cancer Imaging Archive (2017). [Google Scholar]
  • [28].Litjens G, Debats O, Barentsz J, Karssemeijer N, and Huisman H, “Computer-aided detection of prostate cancer in mri,” IEEE Transactions on Medical Imaging 33(5), 1083–1092 (2014). [DOI] [PubMed] [Google Scholar]
  • [29].Cuocolo R, Stanzione A, Castaldo A, De Lucia DR, and Imbriaco M, “Quality control and whole-gland, zonal and lesion annotations for the PROSTATEx challenge public dataset,” European Journal of Radiology 138, 109647 (2021). [DOI] [PubMed] [Google Scholar]
  • [30].van Griethuysen JJ, Fedorov A, Parmar C, Hosny A, Aucoin N, Narayan V, Beets-Tan RG, Fillion-Robin J-C, Pieper S, and Aerts HJ, “Computational radiomics system to decode the radiographic phenotype,” Cancer Research 77(21), e104–e107 (2017). [DOI] [PMC free article] [PubMed] [Google Scholar]
  • [31].Bergholm V, Izaac J, Schuld M, Gogolin C, Ahmed S, Ajith V, Alam MS, Alonso-Linaje G, AkashNarayanan B, Asadi A, et al. , “Pennylane: Automatic differentiation of hybrid quantum-classical computations,” arXiv preprint arXiv:1811.04968 (2018). [Google Scholar]
  • [32].Wang S, Safari M, Li Q, Chang C-W, Qiu RL, Roper J, Yu DS, and Yang X, “Triad: Vision foundation model for 3d magnetic resonance imaging,” Research Square, rs–3 (2025). [Google Scholar]

RESOURCES