Hybrid Metaheuristic Feature Selection for Enhanced Breast Cancer Detection in Digital Mammography: A Radiomics and Deep Learning Approach with Cross-Dataset Validation

Hybrid Metaheuristic Feature Selection for Enhanced Breast Cancer Detection in Digital Mammography: A Radiomics and Deep Learning Approach with Cross-Dataset Validation | Research Square window.SnipcartSettings = { analytics: { enabled: false } }; (function() { var accessVector = localStorage.getItem('access_vector') || ''; window.dataLayer = window.dataLayer || []; if (accessVector) { window.dataLayer.push({ user: { profile: { profileInfo: { snid: accessVector } } } }); } })(); (function(w,d,s,l,i){w[l]=w[l]||[];w[l].push({'gtm.start':new Date().getTime(),event:'gtm.js'});var f=d.getElementsByTagName(s)[0],j=d.createElement(s),dl=l!='dataLayer'?'&l='+l:'';j.async=true;j.src='https://www.googletagmanager.com/gtm.js?id='+i+dl;f.parentNode.insertBefore(j,f);})(window,document,'script','dataLayer','GTM-K279D39R'); Browse Preprints In Review Journals COVID-19 Preprints AJE Video Bytes Research Tools Research Promotion AJE Professional Editing AJE Rubriq About Preprint Platform In Review Editorial Policies Our Team Advisory Board Help Center Sign In Submit a Preprint Cite Share Download PDF Research Article Hybrid Metaheuristic Feature Selection for Enhanced Breast Cancer Detection in Digital Mammography: A Radiomics and Deep Learning Approach with Cross-Dataset Validation Bandar Alshreef This is a preprint; it has not been peer reviewed by a journal. https://doi.org/ 10.21203/rs.3.rs-9239930/v1 This work is licensed under a CC BY 4.0 License Status: Posted Version 1 posted You are reading this latest preprint version Abstract Background Artificial intelligence (AI) shows promise for improving breast cancer detection in mammography, but generalizability across datasets and imaging conditions remains a major challenge. We developed a hybrid metaheuristic feature-selection framework that combines radiomics and deep learning features and evaluated it on a real pilot dataset and a controlled synthetic comparison. Methods A diagnostic model was developed using the public CBIS-DDSM dataset. The framework combined 2,051 IBSI-compliant radiomic features and 2,048-dimensional deep features from a pretrained EfficientNet-B5 model. A hybrid Grasshopper Optimization Algorithm and Crow Search Algorithm (GOA-CSA) was used to select an optimal feature subset for an MLP classifier. A controlled synthetic comparison (N = 16, D = 1114) compared an inventive multi-constraint fitness function against a legacy fitness under collapse-prone conditions. Results On a CBIS-DDSM pilot subset (n = 22, 5-fold cross-validation), the hybrid GOA-CSA model achieved an AUC of 0.858 while reducing the feature count by 95% to 102 features, compared with an all-features baseline AUC of 0.825. In the synthetic comparison, the inventive fitness achieved AUC 0.810 and sensitivity 0.571 versus 0.476 and 0.286 for the legacy fitness. The collapse-prevention mechanism was implemented but was not triggered in this synthetic run, as both models maintained sensitivity greater than zero. Conclusions The hybrid metaheuristic framework improved feature selection performance on both the real pilot and synthetic comparison. The synthetic experiment supports the value of the multi-constraint fitness design, but real-data validation of collapse prevention remains necessary. Breast cancer artificial intelligence digital mammography radiomics deep learning feature selection metaheuristic algorithms decision support Figures Figure 1 Figure 2 Background Breast cancer remains a major global health challenge and one of the leading causes of cancer-related morbidity and mortality among women worldwide. Population-based screening with digital mammography remains central to early detection and improved outcomes, but interpretation is demanding and subject to reader variability, fatigue, and false-positive recalls. Artificial intelligence has increasingly been proposed as a tool to augment mammography screening. Large prospective and implementation studies have suggested that AI-assisted workflows can improve detection performance and reduce radiologist workload. However, robust performance across heterogeneous clinical datasets remains difficult to achieve. Two major paradigms in breast imaging AI are radiomics and deep learning. Radiomics captures quantitative image descriptors such as texture, shape, and intensity patterns, while deep learning models learn higher-level hierarchical representations directly from image data. Both have shown promise, but both also face limitations related to robustness, interpretability, overfitting, and cross-dataset generalizability. Feature selection is especially important when radiomic and deep features are combined into a high-dimensional representation with limited sample size. Metaheuristic optimization algorithms such as the Grasshopper Optimization Algorithm (GOA) and Crow Search Algorithm (CSA) have been used to explore large feature spaces, but their performance can be unstable under low-sample, imbalanced conditions. This study therefore aimed to develop and evaluate a hybrid metaheuristic feature-selection framework for digital mammography that combines radiomics and deep learning features, evaluates performance on a real CBIS-DDSM pilot subset, and examines a controlled synthetic comparison designed to test the behavior of a multi-constraint fitness function under collapse-prone conditions. Methods Datasets This retrospective computational study used two publicly available, anonymized datasets: the Curated Breast Imaging Subset of DDSM (CBIS-DDSM) for model development and internal pilot evaluation, and VinDr-Mammo as the planned external validation dataset. For the pilot experiment reported here, a subset of 22 lesion regions of interest from CBIS-DDSM (10 benign and 12 malignant) was used. The VinDr-Mammo dataset was selected for future external validation because it represents an independent, large-scale full-field digital mammography cohort acquired under different clinical conditions. Preprocessing A standardized preprocessing workflow was applied to prepare the images for analysis. The pipeline included consistent handling of image polarity, normalization of pixel intensities, and resizing of regions of interest to a common spatial format suitable for feature extraction and model input. Where available, lesion segmentation masks were used to define regions of interest. For datasets without lesion masks, regions of interest were defined from available lesion localization annotations. Feature extraction A multi-modal feature representation was constructed by combining handcrafted radiomic features with deep features derived from a pretrained convolutional neural network. Radiomic features were extracted with pyradiomics using an IBSI-aligned parameter configuration and included first-order, shape-based, and texture-derived descriptors, along with transformed-image features where applicable. Deep features were extracted using a pretrained EfficientNet-B5 model. The output of the global average pooling layer was used as a 2,048-dimensional representation for each region of interest. After concatenation, an initial filtering stage removed low-variance or otherwise unsuitable features before optimization. Hybrid GOA-CSA feature selection A hybrid metaheuristic feature-selection procedure combining GOA and CSA was used to search for optimized subsets of radiomic and deep features. In the real-data pilot, the hybrid GOA-CSA model using the inventive fitness selected 102 features from 2,051 candidate features, corresponding to approximately 95% reduction. For the synthetic comparison, a multi-constraint fitness design was evaluated against a legacy fitness. The synthetic dataset contained N = 16 samples and D = 1114 features, with strong signal features, weak collapse-trap features, a correlated block, and noise dimensions designed to mimic instability and collapse-prone conditions. Classification and evaluation A multilayer perceptron (MLP) classifier was used to evaluate selected feature subsets. Performance was assessed with 5-fold stratified cross-validation for both the real CBIS-DDSM pilot and the synthetic comparative experiment to maintain consistency across evaluations. The primary performance metrics were area under the receiver operating characteristic curve (AUC), sensitivity, specificity, and F1-score. For the synthetic experiment, stability was additionally summarized using mean pairwise Jaccard similarity across selected feature subsets. Results Pilot experiment on real CBIS-DDSM data (n = 22) In the CBIS-DDSM pilot subset (n = 22), the all-features baseline achieved an AUC of 0.825, accuracy of 0.818, sensitivity of 0.667, specificity of 1.000, and F1-score of 0.800. The hybrid GOA-CSA model selected 102 features and achieved an AUC of 0.858, accuracy of 0.727, sensitivity of 0.833, specificity of 0.600, and F1-score of 0.769. Although specificity decreased, the feature-selected model achieved higher sensitivity, which is clinically important in screening-oriented detection tasks. Detailed model performance metrics are summarized in Table 1 , and the corresponding ROC curves are shown in Fig. 1 . [Insert Table about here] [Insert Fig. 1 about here] Table 1 Model performance on the CBIS-DDSM pilot cohort (n = 22, 5-fold stratified cross-validation). Model Features AUC Acc. Sens. Spec. F1-score PPV NPV Baseline 2,051 0.825 0.818 0.667 1.000 0.800 1.000 0.714 Hybrid GOA-CSA 102 0.858 0.727 0.833 0.600 0.769 0.714 0.750 Comparative experiment: legacy vs inventive fitness on synthetic data The synthetic comparison (N = 16, D = 1114) directly compared a legacy fitness with the inventive multi-constraint fitness. The inventive fitness achieved higher AUC (0.810 vs 0.476), higher sensitivity (0.571 vs 0.286), and higher F1-score (0.727 vs 0.444), while both models retained specificity of 1.000. The collapse-prevention penalty was implemented but was not triggered in this run because neither model reached zero sensitivity. Accordingly, this experiment should be interpreted as evidence that the inventive fitness improved performance in a controlled collapse-prone setting, not as proof that collapse prevention has already been demonstrated on real mammography data. Comparative synthetic performance metrics are summarized in Table 2 , and the comparative performance panels are shown in Fig. 2 . [Insert Table 2 about here] [Insert Fig. 2 about here] The collapse trap itself was validated: when evaluated using only trap features, an unbalanced logistic regression produced sensitivity of 0.0, confirming that the synthetic design contained a genuine failure mode. Table 2 Legacy vs inventive fitness on synthetic data (5-fold cross-validation, MLP evaluation). Model Features selected AUC Sensitivity Specificity Stability (Jaccard) Collapsed? Legacy 618 0.476 0.286 1.000 0.489 No Inventive 639 0.810 0.571 1.000 0.481 No Planned full-scale validation Full-scale experiments on the complete CBIS-DDSM training cohort (1,566 patients) and planned external validation on the VinDr-Mammo dataset (5,000 patients) will be conducted following completion of dataset download, harmonization, and feature extraction. Discussion Synthetic evidence of the inventive fitness The controlled synthetic comparison provides direct evidence that the inventive multi-constraint fitness function outperformed the legacy fitness in a collapse-prone setting. The improvement in discrimination and sensitivity suggests that the balanced evaluation and stability-aware design better reward feature subsets that retain clinically relevant minority-class signal. However, the collapse-prevention penalty was not triggered in this run. Therefore, the synthetic evidence supports the value of the fitness design but does not yet prove active collapse prevention under real-world mammography conditions. Real-data pilot The real CBIS-DDSM pilot demonstrated that the hybrid GOA-CSA approach can produce a compact feature set with improved AUC and sensitivity relative to the all-features baseline. This suggests that joint optimization over radiomic and deep features can improve performance even on a small pilot cohort. At the same time, the real-data pilot was not designed as a direct collapse-prevention test. It should therefore be interpreted as a preliminary real-data performance study rather than as confirmation that the inventive fitness prevents collapse in practice. Remaining validation The most important remaining step is direct real-data replication of the previously documented collapse scenario using a legacy fitness on CBIS-DDSM, followed by demonstration that the inventive fitness avoids that failure mode on the same real-data setting. External validation on VinDr-Mammo also remains essential to evaluate cross-dataset generalizability. Limitations This study has several limitations. First, the real-data pilot was small (n = 22), which limits statistical confidence and generalizability. Second, the study was retrospective and computational in nature, without prospective workflow testing. Third, the synthetic comparison, although intentionally designed to reproduce collapse-prone conditions, remains a controlled experiment rather than a substitute for real-data validation. Finally, external validation on VinDr-Mammo is planned but not yet completed. Future work Future work will focus on full-scale internal validation on the complete CBIS-DDSM cohort, external validation on VinDr-Mammo, and direct testing of collapse prevention on real data. Additional work should also assess robustness across acquisition settings and explore model interpretability and deployment considerations relevant to clinical decision support. Conclusions This study presents a hybrid metaheuristic feature-selection framework for digital mammography that combines radiomics and deep learning features within a GOA-CSA optimization pipeline. On a real CBIS-DDSM pilot, the framework improved AUC and sensitivity while substantially reducing feature count. In a controlled synthetic comparison, the inventive multi-constraint fitness outperformed a legacy fitness in AUC and sensitivity. The collapse-prevention mechanism was implemented but was not triggered in this synthetic run, and real-data confirmation remains necessary before stronger claims about collapse prevention can be made. Abbreviations AI Artificial Intelligence AUC Area Under the ROC Curve CBIS-DDSM Curated Breast Imaging Subset of DDSM CSA Crow Search Algorithm CV Cross-Validation DDSM Digital Database for Screening Mammography GOA Grasshopper Optimization Algorithm IBSI Image Biomarker Standardization Initiative MLP Multilayer Perceptron NPV Negative Predictive Value PPV Positive Predictive Value ROC Receiver Operating Characteristic. Declarations Ethics approval and consent to participate Not applicable. This study used publicly available de-identified imaging datasets and synthetic data; no direct patient contact or intervention occurred. Consent for publication Not applicable. Competing interests The author declares that there are no competing interests. Funding: No specific funding was received for this work. Author Contribution The author conceptualized the study, developed the methodology, curated the computational workflow, interpreted the results, and drafted the manuscript. Acknowledgement The author would like to thank the Deanship of Scientific Research at Shaqra University for supporting this work. The author also acknowledges the providers of the publicly available CBIS-DDSM and VinDr-Mammo datasets used in this study. Data Availability CBIS-DDSM and VinDr-Mammo are publicly available from their original repositories. The synthetic comparative experiment is generated from code included in the project repository. The code for feature extraction, feature selection, model training, and evaluation is available at https://github.com/bsalshreef/Hybrid-Metaheuristic-Mammography-AI. References Chang YW et al. Artificial intelligence for breast cancer screening in mammography (AI-STREAM): preliminary analysis of a prospective multicenter cohort study. Nat Commun. 2025;[Epub ahead of print]. Hernström V et al. Screening performance and characteristics of breast artificial intelligence in population-based mammography programmes. Lancet Digit Health. 2025;[Epub ahead of print]. Lång K, et al. AI-supported breast cancer screening—new results suggest even higher accuracy. Lund University report summarizing MASAI trial results; 2026. Dembrower K et al. Nation-wide real-world implementation of AI for cancer detection in population-based mammography screening. Nat Med. 2025;[Epub ahead of print]. Radiological Society of North America. AI identifies interval breast cancer. RSNA News. 2025 Oct 27. Goh S, et al. Challenges in implementing artificial intelligence in breast cancer screening programs: qualitative insights from radiologists and administrators. J Med Internet Res. 2025;27:e62941. Breastcancer.org. Using AI (Artificial Intelligence) to Detect Breast Cancer. 2024 Mar 25. Available from: https://www.breastcancer.org/ Hadjievskyi N, et al. Robust radiomic feature selection in digital mammography: understanding the effect of imaging acquisition physics using phantom and clinical data analysis. Cancers (Basel). 2020;12(3):700. Gillies RJ, Kinahan PE, Hricak H. Radiomics: images are more than pictures, they are data. Radiology. 2016;278(2):563–77. Alsaadi T et al. MRI-based machine learning radiomics for prediction of breast cancer molecular subtypes. Diagn Interv Imaging. 2024;[Epub ahead of print]. Li X, et al. Advances in sampling, feature selection, and multi-omics integration for radiomics-based cancer prediction. Cancers (Basel). 2025;17(9):1610. Banu A, et al. Hybrid optimization algorithm based feature selection for mammogram images and detecting the breast mass using multilayer perceptron classifier. Comput Intell. 2022;38(3):888–910. Ahmed S, et al. Optimizing cancer classification: a metaheuristic-driven review of feature selection and deep neural network optimization (2012–2025). Technol Cancer Res Treat. 2025;24:15330338251375817. Kim H, et al. Artificial intelligence in breast ultrasound: a systematic review of clinical performance, bias, and implementation barriers. Front Oncol. 2025;15:1619364. Jafari M, et al. Breast cancer prediction using hybrid machine learning and nature-inspired adaptive optimization algorithms. J Contemp Health Res. 2023;3(4):1–10. Erickson BJ, Korfiatis P, Akkus Z, Kline TL. Radiology artificial intelligence: radiological images and machine learning. Radiographics. 2019;39(7):1923–39. Park SH, Han K. Methodologic guide for evaluating clinical performance and effect of artificial intelligence technology for medical diagnosis and prediction. Radiology. 2018;286(3):800–9. Mongan J, Moy L, Kahn CE Jr. Checklist for artificial intelligence in medical imaging (CLAIM): a guide for authors and reviewers. Radiol Artif Intell. 2020;2(2):e200029. Lambin P, et al. Radiomics: the bridge between medical imaging and personalized medicine. Nat Rev Clin Oncol. 2017;14(12):749–62. Yala A, et al. Toward robust mammography-based models for breast cancer risk. Sci Transl Med. 2019;11(575):eaax6883. Truhn D, et al. Radiomics in mammography: evaluation of feature stability and acquisition variability. Eur Radiol. 2021;31(5):3216–27. Pesapane F, Rotili A, Penco S, Napolitano A, Houssami N, Sardanelli F. Artificial intelligence and mammography: where are we now? Eur J Radiol Exp. 2022;6(1):1–18. Chae EY, et al. Deep learning for breast cancer detection on digital mammograms: systematic review and meta-analysis. Eur J Radiol. 2023;161:110749. Sechopoulos I, Mann RM. Stand-alone artificial intelligence for breast cancer detection in mammography: current evidence and real-world considerations. Nat Rev Clin Oncol. 2023;20(9):540–56. Dalmis MU, et al. Artificial intelligence for reducing workload in breast cancer screening with mammography: overview of current evidence. Eur J Radiol. 2024;165:111129. Houssami N et al. Will AI reshape breast cancer screening? ESMO Daily Reporter. 2025 May 11. Litjens G, et al. A survey on deep learning in medical image analysis. Med Image Anal. 2017;42:60–88. van Griethuysen JJM, et al. Computational radiomics system to decode the radiographic phenotype. Cancer Res. 2017;77(21):e104–7. Lee RS, Gimenez F, Hoogi A, et al. A curated mammography data set for use in computer-aided detection and diagnosis research. Sci Data. 2017;4:170177. Clark K, Vendt B, Smith K, et al. The Cancer Imaging Archive (TCIA): maintaining and operating a public information repository. J Digit Imaging. 2013;26(6):1045–57. Nguyen T, et al. VinDr-Mammo: a large-scale benchmark dataset for computer-aided diagnosis in full-field digital mammography. Sci Data. 2023;10:262. Al-Helo S, et al. Narrative review of open-access mammography databases for artificial intelligence research. High Yield Med Rev. 2024;4(1):e24. Larson DB, et al. Ethics of using and sharing clinical imaging data for research and artificial intelligence development. Radiology. 2020;295(3):E131–40. Ethical design of. data-driven decision support tools for breast cancer screening. JMIR Cancer. 2025;e65566. Full Field Digital. Mammography Dataset from a Population-based Screening Program. Sci Data. 2025. Zhou B, et al. Immunity-based Ebola optimization search algorithm for cancer classification using gene expression datasets. Sci Rep. 2022;12:18056. Youn I, et al. External validation of AI algorithms for breast cancer detection in digital mammography. Radiology. 2024;312(1):56–67. Rodriguez-Ruiz A, et al. Stand-alone artificial intelligence for breast cancer detection in screening mammography: comparison with 101 radiologists. J Natl Cancer Inst. 2019;111(9):916–22. McKinney SM, et al. International evaluation of an AI system for breast cancer screening. Nature. 2020;577(7788):89–94. Yala A, Lehman CD, Schuster T, Portnoi T, Barzilay R. A deep learning mammography-based model for improved breast cancer risk prediction. Radiology. 2019;292(1):60–6. Duffy SW, et al. Randomized trials of mammography screening: updated overview of long-term mortality benefits. Radiology. 2020;296(2):221–7. Mann RM, Hooley RJ, Barr RG. Breast ultrasound in 2025 and beyond: impact of artificial intelligence and radiomics. AJR Am J Roentgenol. 2025;[Epub ahead of print]. Tan M, Le Q. EfficientNet: rethinking model scaling for convolutional neural networks. Proc ICML. 2019;6105–14. Additional Declarations No competing interests reported. Cite Share Download PDF Status: Posted Version 1 posted You are reading this latest preprint version Research Square lets you share your work early, gain feedback from the community, and start making changes to your manuscript prior to peer review in a journal. As a division of Research Square Company, we’re committed to making research communication faster, fairer, and more useful. We do this by developing innovative software and high quality services for the global research community. Our growing team is made up of researchers and industry professionals working together to solve the most critical problems facing scientific publishing. Also discoverable on Platform About Our Team In Review Editorial Policies Advisory Board Help Center Resources Author Services Accessibility API Access RSS feed Manage Cookie Preferences © Research Square 2026 | ISSN 2693-5015 (online) Privacy Policy Terms of Service Do Not Sell My Personal Information {"props":{"pageProps":{"initialData":{"identity":"rs-9239930","acceptedTermsAndConditions":true,"allowDirectSubmit":true,"archivedVersions":[],"articleType":"Research Article","associatedPublications":[],"authors":[{"id":614612795,"identity":"afc1ce7f-ab41-4c53-9375-5d7efd483fa2","order_by":0,"name":"Bandar Alshreef","email":"data:image/png;base64,iVBORw0KGgoAAAANSUhEUgAAAZAAAAAyAQMAAABI0h/eAAAABlBMVEX///8AAABVwtN+AAAACXBIWXMAAA7EAAAOxAGVKw4bAAAA7ElEQVRIiWNgGAWjYBACAwYGxgOMDUAW8+EDIAEwm5AWBogWtrQEkrXkGBCnxZz/8IODP3fck+Nn4/n8mYfBRnbDAfaHH/BpsZyRZnCY90yxsWQb7zZpHoY04w0HeIwl8DrsBoPBYca2hMQN93u3MfMwHE4EamHAr+X88Q8Hf7Yl1O8/xvMY6LD/QC3sj3/g1XIgx+AAb1tCggEbDwPQYQeAWhjM8NpiOSOn4DBQi+GMY2xmknMMko1nHuYxs8CnxZz/+MaHQIfJ87cxP/7wpsJOtu94++Mb+LSguxOImUlQPwpGwSgYBaMAOwAAUUtONoi+Vw8AAAAASUVORK5CYII=","orcid":"","institution":"Shaqra University","correspondingAuthor":true,"prefix":"","firstName":"Bandar","middleName":"","lastName":"Alshreef","suffix":""}],"badges":[],"createdAt":"2026-03-27 04:23:47","currentVersionCode":1,"declarations":"","doi":"10.21203/rs.3.rs-9239930/v1","doiUrl":"https://doi.org/10.21203/rs.3.rs-9239930/v1","draftVersion":[],"editorialEvents":[],"editorialNote":"","failedWorkflow":false,"files":[{"id":106096140,"identity":"f7fc2cb4-97e4-42c7-b1dc-a73ec25ee428","added_by":"auto","created_at":"2026-04-03 11:53:00","extension":"png","order_by":1,"title":"Figure 1","display":"","copyAsset":false,"role":"figure","size":305658,"visible":true,"origin":"","legend":"\u003cp\u003eReceiver operating characteristic (ROC) curves for the all-features baseline and hybrid GOA-CSA models on the CBIS-DDSM internal validation cohort (5-fold stratified cross-validation, n=22). Hybrid GOA-CSA (AUC = 0.858) outperformed the all-features baseline (AUC = 0.825).\u003c/p\u003e","description":"","filename":"floatimage1.png","url":"https://assets-eu.researchsquare.com/files/rs-9239930/v1/ca788105aabd0ded7d22534d.png"},{"id":106096164,"identity":"eab28c1d-ff2b-4aca-b798-ce17e3b817da","added_by":"auto","created_at":"2026-04-03 11:53:14","extension":"png","order_by":2,"title":"Figure 2","display":"","copyAsset":false,"role":"figure","size":470315,"visible":true,"origin":"","legend":"\u003cp\u003eComparative performance of legacy vs inventive fitness on synthetic data. Panel A shows ROC curves. Panel B summarizes classification metrics. Panel C summarizes feature quality and stability. Panels D–F provide convergence history, confusion matrices, and a summary table.\u003c/p\u003e","description":"","filename":"floatimage2.png","url":"https://assets-eu.researchsquare.com/files/rs-9239930/v1/7b47bd023608ceec66be534b.png"},{"id":108180742,"identity":"0522c22b-ba23-43a6-8786-5cc7838a6e86","added_by":"auto","created_at":"2026-04-30 08:52:55","extension":"pdf","order_by":0,"title":"","display":"","copyAsset":false,"role":"manuscript-pdf","size":815136,"visible":true,"origin":"","legend":"","description":"","filename":"manuscript.pdf","url":"https://assets-eu.researchsquare.com/files/rs-9239930/v1/7bb4745d-20c2-4ed2-a310-29d462ce1ad0.pdf"}],"financialInterests":"No competing interests reported.","formattedTitle":"Hybrid Metaheuristic Feature Selection for Enhanced Breast Cancer Detection in Digital Mammography: A Radiomics and Deep Learning Approach with Cross-Dataset Validation","fulltext":[{"header":"Background","content":"\u003cp\u003eBreast cancer remains a major global health challenge and one of the leading causes of cancer-related morbidity and mortality among women worldwide. Population-based screening with digital mammography remains central to early detection and improved outcomes, but interpretation is demanding and subject to reader variability, fatigue, and false-positive recalls.\u003c/p\u003e \u003cp\u003eArtificial intelligence has increasingly been proposed as a tool to augment mammography screening. Large prospective and implementation studies have suggested that AI-assisted workflows can improve detection performance and reduce radiologist workload. However, robust performance across heterogeneous clinical datasets remains difficult to achieve.\u003c/p\u003e \u003cp\u003eTwo major paradigms in breast imaging AI are radiomics and deep learning. Radiomics captures quantitative image descriptors such as texture, shape, and intensity patterns, while deep learning models learn higher-level hierarchical representations directly from image data. Both have shown promise, but both also face limitations related to robustness, interpretability, overfitting, and cross-dataset generalizability.\u003c/p\u003e \u003cp\u003eFeature selection is especially important when radiomic and deep features are combined into a high-dimensional representation with limited sample size. Metaheuristic optimization algorithms such as the Grasshopper Optimization Algorithm (GOA) and Crow Search Algorithm (CSA) have been used to explore large feature spaces, but their performance can be unstable under low-sample, imbalanced conditions.\u003c/p\u003e \u003cp\u003eThis study therefore aimed to develop and evaluate a hybrid metaheuristic feature-selection framework for digital mammography that combines radiomics and deep learning features, evaluates performance on a real CBIS-DDSM pilot subset, and examines a controlled synthetic comparison designed to test the behavior of a multi-constraint fitness function under collapse-prone conditions.\u003c/p\u003e"},{"header":"Methods","content":"\u003cdiv id=\"Sec3\" class=\"Section2\"\u003e \u003ch2\u003eDatasets\u003c/h2\u003e \u003cp\u003eThis retrospective computational study used two publicly available, anonymized datasets: the Curated Breast Imaging Subset of DDSM (CBIS-DDSM) for model development and internal pilot evaluation, and VinDr-Mammo as the planned external validation dataset.\u003c/p\u003e \u003cp\u003eFor the pilot experiment reported here, a subset of 22 lesion regions of interest from CBIS-DDSM (10 benign and 12 malignant) was used. The VinDr-Mammo dataset was selected for future external validation because it represents an independent, large-scale full-field digital mammography cohort acquired under different clinical conditions.\u003c/p\u003e \u003c/div\u003e\n\u003ch3\u003ePreprocessing\u003c/h3\u003e\n\u003cp\u003eA standardized preprocessing workflow was applied to prepare the images for analysis. The pipeline included consistent handling of image polarity, normalization of pixel intensities, and resizing of regions of interest to a common spatial format suitable for feature extraction and model input.\u003c/p\u003e \u003cp\u003eWhere available, lesion segmentation masks were used to define regions of interest. For datasets without lesion masks, regions of interest were defined from available lesion localization annotations.\u003c/p\u003e\n\u003ch3\u003eFeature extraction\u003c/h3\u003e\n\u003cp\u003eA multi-modal feature representation was constructed by combining handcrafted radiomic features with deep features derived from a pretrained convolutional neural network.\u003c/p\u003e \u003cp\u003eRadiomic features were extracted with pyradiomics using an IBSI-aligned parameter configuration and included first-order, shape-based, and texture-derived descriptors, along with transformed-image features where applicable.\u003c/p\u003e \u003cp\u003eDeep features were extracted using a pretrained EfficientNet-B5 model. The output of the global average pooling layer was used as a 2,048-dimensional representation for each region of interest. After concatenation, an initial filtering stage removed low-variance or otherwise unsuitable features before optimization.\u003c/p\u003e\n\u003ch3\u003eHybrid GOA-CSA feature selection\u003c/h3\u003e\n\u003cp\u003eA hybrid metaheuristic feature-selection procedure combining GOA and CSA was used to search for optimized subsets of radiomic and deep features.\u003c/p\u003e \u003cp\u003eIn the real-data pilot, the hybrid GOA-CSA model using the inventive fitness selected 102 features from 2,051 candidate features, corresponding to approximately 95% reduction.\u003c/p\u003e \u003cp\u003eFor the synthetic comparison, a multi-constraint fitness design was evaluated against a legacy fitness. The synthetic dataset contained N\u0026thinsp;=\u0026thinsp;16 samples and D\u0026thinsp;=\u0026thinsp;1114 features, with strong signal features, weak collapse-trap features, a correlated block, and noise dimensions designed to mimic instability and collapse-prone conditions.\u003c/p\u003e\n\u003ch3\u003eClassification and evaluation\u003c/h3\u003e\n\u003cp\u003eA multilayer perceptron (MLP) classifier was used to evaluate selected feature subsets. Performance was assessed with 5-fold stratified cross-validation for both the real CBIS-DDSM pilot and the synthetic comparative experiment to maintain consistency across evaluations.\u003c/p\u003e \u003cp\u003eThe primary performance metrics were area under the receiver operating characteristic curve (AUC), sensitivity, specificity, and F1-score. For the synthetic experiment, stability was additionally summarized using mean pairwise Jaccard similarity across selected feature subsets.\u003c/p\u003e"},{"header":"Results","content":"\u003cdiv id=\"Sec9\" class=\"Section2\"\u003e \u003ch2\u003ePilot experiment on real CBIS-DDSM data (n\u0026thinsp;=\u0026thinsp;22)\u003c/h2\u003e \u003cp\u003eIn the CBIS-DDSM pilot subset (n\u0026thinsp;=\u0026thinsp;22), the all-features baseline achieved an AUC of 0.825, accuracy of 0.818, sensitivity of 0.667, specificity of 1.000, and F1-score of 0.800. The hybrid GOA-CSA model selected 102 features and achieved an AUC of 0.858, accuracy of 0.727, sensitivity of 0.833, specificity of 0.600, and F1-score of 0.769. Although specificity decreased, the feature-selected model achieved higher sensitivity, which is clinically important in screening-oriented detection tasks.\u003c/p\u003e \u003cp\u003eDetailed model performance metrics are summarized in Table\u0026nbsp;\u003cspan refid=\"Tab1\" class=\"InternalRef\"\u003e1\u003c/span\u003e, and the corresponding ROC curves are shown in Fig.\u0026nbsp;\u003cspan refid=\"Fig1\" class=\"InternalRef\"\u003e1\u003c/span\u003e.\u003c/p\u003e \u003cp\u003e \u003c/p\u003e \u003c/div\u003e\n\u003ch3\u003e[Insert Table about here]\u003c/h3\u003e\n\u003cdiv id=\"Sec11\" class=\"Section2\"\u003e \u003ch2\u003e[Insert Fig.\u0026nbsp;\u003cspan refid=\"Fig1\" class=\"InternalRef\"\u003e1\u003c/span\u003e about here]\u003c/h2\u003e \u003cp\u003e \u003cdiv class=\"gridtable\"\u003e\u003ctable float=\"Yes\" id=\"Tab1\" border=\"1\"\u003e \u003ccaption language=\"En\"\u003e \u003cdiv class=\"CaptionNumber\"\u003eTable 1\u003c/div\u003e \u003cdiv class=\"CaptionContent\"\u003e \u003cp\u003eModel performance on the CBIS-DDSM pilot cohort (n\u0026thinsp;=\u0026thinsp;22, 5-fold stratified cross-validation).\u003c/p\u003e \u003c/div\u003e \u003c/caption\u003e \u003ccolgroup cols=\"9\"\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c1\" colnum=\"1\"\u003e\u003c/div\u003e \u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c2\" colnum=\"2\"\u003e\u003c/div\u003e \u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c3\" colnum=\"3\"\u003e\u003c/div\u003e \u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c4\" colnum=\"4\"\u003e\u003c/div\u003e \u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c5\" colnum=\"5\"\u003e\u003c/div\u003e \u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c6\" colnum=\"6\"\u003e\u003c/div\u003e \u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c7\" colnum=\"7\"\u003e\u003c/div\u003e \u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c8\" colnum=\"8\"\u003e\u003c/div\u003e \u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c9\" colnum=\"9\"\u003e\u003c/div\u003e \u003cthead\u003e \u003ctr\u003e \u003cth align=\"left\" colname=\"c1\"\u003e \u003cp\u003eModel\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c2\"\u003e \u003cp\u003eFeatures\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c3\"\u003e \u003cp\u003eAUC\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c4\"\u003e \u003cp\u003eAcc.\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c5\"\u003e \u003cp\u003eSens.\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c6\"\u003e \u003cp\u003eSpec.\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c7\"\u003e \u003cp\u003eF1-score\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c8\"\u003e \u003cp\u003ePPV\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c9\"\u003e \u003cp\u003eNPV\u003c/p\u003e \u003c/th\u003e \u003c/tr\u003e \u003c/thead\u003e \u003ctbody\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eBaseline\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e2,051\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e0.825\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e0.818\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e0.667\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c6\"\u003e \u003cp\u003e1.000\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c7\"\u003e \u003cp\u003e0.800\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c8\"\u003e \u003cp\u003e1.000\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c9\"\u003e \u003cp\u003e0.714\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eHybrid GOA-CSA\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e102\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e0.858\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e0.727\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e0.833\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c6\"\u003e \u003cp\u003e0.600\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c7\"\u003e \u003cp\u003e0.769\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c8\"\u003e \u003cp\u003e0.714\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c9\"\u003e \u003cp\u003e0.750\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003c/tbody\u003e \u003c/colgroup\u003e \u003c/table\u003e\u003c/div\u003e \u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec12\" class=\"Section2\"\u003e \u003ch2\u003eComparative experiment: legacy vs inventive fitness on synthetic data\u003c/h2\u003e \u003cp\u003eThe synthetic comparison (N\u0026thinsp;=\u0026thinsp;16, D\u0026thinsp;=\u0026thinsp;1114) directly compared a legacy fitness with the inventive multi-constraint fitness. The inventive fitness achieved higher AUC (0.810 vs 0.476), higher sensitivity (0.571 vs 0.286), and higher F1-score (0.727 vs 0.444), while both models retained specificity of 1.000.\u003c/p\u003e \u003cp\u003eThe collapse-prevention penalty was implemented but was not triggered in this run because neither model reached zero sensitivity. Accordingly, this experiment should be interpreted as evidence that the inventive fitness improved performance in a controlled collapse-prone setting, not as proof that collapse prevention has already been demonstrated on real mammography data.\u003c/p\u003e \u003cp\u003eComparative synthetic performance metrics are summarized in Table\u0026nbsp;\u003cspan refid=\"Tab2\" class=\"InternalRef\"\u003e2\u003c/span\u003e, and the comparative performance panels are shown in Fig.\u0026nbsp;\u003cspan refid=\"Fig2\" class=\"InternalRef\"\u003e2\u003c/span\u003e.\u003c/p\u003e \u003cp\u003e \u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec13\" class=\"Section2\"\u003e \u003ch2\u003e[Insert Table\u0026nbsp;\u003cspan refid=\"Tab2\" class=\"InternalRef\"\u003e2\u003c/span\u003e about here]\u003c/h2\u003e \u003cdiv id=\"Sec14\" class=\"Section3\"\u003e \u003ch2\u003e[Insert Fig.\u0026nbsp;\u003cspan refid=\"Fig2\" class=\"InternalRef\"\u003e2\u003c/span\u003e about here]\u003c/h2\u003e \u003cp\u003eThe collapse trap itself was validated: when evaluated using only trap features, an unbalanced logistic regression produced sensitivity of 0.0, confirming that the synthetic design contained a genuine failure mode.\u003c/p\u003e \u003cp\u003e \u003cdiv class=\"gridtable\"\u003e\u003ctable float=\"Yes\" id=\"Tab2\" border=\"1\"\u003e \u003ccaption language=\"En\"\u003e \u003cdiv class=\"CaptionNumber\"\u003eTable 2\u003c/div\u003e \u003cdiv class=\"CaptionContent\"\u003e \u003cp\u003eLegacy vs inventive fitness on synthetic data (5-fold cross-validation, MLP evaluation).\u003c/p\u003e \u003c/div\u003e \u003c/caption\u003e \u003ccolgroup cols=\"7\"\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c1\" colnum=\"1\"\u003e\u003c/div\u003e \u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c2\" colnum=\"2\"\u003e\u003c/div\u003e \u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c3\" colnum=\"3\"\u003e\u003c/div\u003e \u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c4\" colnum=\"4\"\u003e\u003c/div\u003e \u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c5\" colnum=\"5\"\u003e\u003c/div\u003e \u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c6\" colnum=\"6\"\u003e\u003c/div\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c7\" colnum=\"7\"\u003e\u003c/div\u003e \u003cthead\u003e \u003ctr\u003e \u003cth align=\"left\" colname=\"c1\"\u003e \u003cp\u003eModel\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c2\"\u003e \u003cp\u003eFeatures selected\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c3\"\u003e \u003cp\u003eAUC\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c4\"\u003e \u003cp\u003eSensitivity\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c5\"\u003e \u003cp\u003eSpecificity\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c6\"\u003e \u003cp\u003eStability (Jaccard)\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c7\"\u003e \u003cp\u003eCollapsed?\u003c/p\u003e \u003c/th\u003e \u003c/tr\u003e \u003c/thead\u003e \u003ctbody\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eLegacy\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e618\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e0.476\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e0.286\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e1.000\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c6\"\u003e \u003cp\u003e0.489\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c7\"\u003e \u003cp\u003eNo\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eInventive\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e639\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e0.810\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e0.571\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e1.000\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c6\"\u003e \u003cp\u003e0.481\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c7\"\u003e \u003cp\u003eNo\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003c/tbody\u003e \u003c/colgroup\u003e \u003c/table\u003e\u003c/div\u003e \u003c/p\u003e \u003c/div\u003e \u003c/div\u003e \u003cdiv id=\"Sec15\" class=\"Section2\"\u003e \u003ch2\u003ePlanned full-scale validation\u003c/h2\u003e \u003cp\u003eFull-scale experiments on the complete CBIS-DDSM training cohort (1,566 patients) and planned external validation on the VinDr-Mammo dataset (5,000 patients) will be conducted following completion of dataset download, harmonization, and feature extraction.\u003c/p\u003e \u003c/div\u003e"},{"header":"Discussion","content":"\u003cdiv id=\"Sec17\" class=\"Section2\"\u003e \u003ch2\u003eSynthetic evidence of the inventive fitness\u003c/h2\u003e \u003cp\u003eThe controlled synthetic comparison provides direct evidence that the inventive multi-constraint fitness function outperformed the legacy fitness in a collapse-prone setting. The improvement in discrimination and sensitivity suggests that the balanced evaluation and stability-aware design better reward feature subsets that retain clinically relevant minority-class signal.\u003c/p\u003e \u003cp\u003eHowever, the collapse-prevention penalty was not triggered in this run. Therefore, the synthetic evidence supports the value of the fitness design but does not yet prove active collapse prevention under real-world mammography conditions.\u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec18\" class=\"Section2\"\u003e \u003ch2\u003eReal-data pilot\u003c/h2\u003e \u003cp\u003eThe real CBIS-DDSM pilot demonstrated that the hybrid GOA-CSA approach can produce a compact feature set with improved AUC and sensitivity relative to the all-features baseline. This suggests that joint optimization over radiomic and deep features can improve performance even on a small pilot cohort.\u003c/p\u003e \u003cp\u003eAt the same time, the real-data pilot was not designed as a direct collapse-prevention test. It should therefore be interpreted as a preliminary real-data performance study rather than as confirmation that the inventive fitness prevents collapse in practice.\u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec19\" class=\"Section2\"\u003e \u003ch2\u003eRemaining validation\u003c/h2\u003e \u003cp\u003eThe most important remaining step is direct real-data replication of the previously documented collapse scenario using a legacy fitness on CBIS-DDSM, followed by demonstration that the inventive fitness avoids that failure mode on the same real-data setting.\u003c/p\u003e \u003cp\u003eExternal validation on VinDr-Mammo also remains essential to evaluate cross-dataset generalizability.\u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec20\" class=\"Section2\"\u003e \u003ch2\u003eLimitations\u003c/h2\u003e \u003cp\u003eThis study has several limitations. First, the real-data pilot was small (n\u0026thinsp;=\u0026thinsp;22), which limits statistical confidence and generalizability. Second, the study was retrospective and computational in nature, without prospective workflow testing. Third, the synthetic comparison, although intentionally designed to reproduce collapse-prone conditions, remains a controlled experiment rather than a substitute for real-data validation. Finally, external validation on VinDr-Mammo is planned but not yet completed.\u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec21\" class=\"Section2\"\u003e \u003ch2\u003eFuture work\u003c/h2\u003e \u003cp\u003eFuture work will focus on full-scale internal validation on the complete CBIS-DDSM cohort, external validation on VinDr-Mammo, and direct testing of collapse prevention on real data. Additional work should also assess robustness across acquisition settings and explore model interpretability and deployment considerations relevant to clinical decision support.\u003c/p\u003e \u003c/div\u003e"},{"header":"Conclusions","content":"\u003cp\u003eThis study presents a hybrid metaheuristic feature-selection framework for digital mammography that combines radiomics and deep learning features within a GOA-CSA optimization pipeline. On a real CBIS-DDSM pilot, the framework improved AUC and sensitivity while substantially reducing feature count. In a controlled synthetic comparison, the inventive multi-constraint fitness outperformed a legacy fitness in AUC and sensitivity. The collapse-prevention mechanism was implemented but was not triggered in this synthetic run, and real-data confirmation remains necessary before stronger claims about collapse prevention can be made.\u003c/p\u003e"},{"header":"Abbreviations","content":"\u003cdiv class=\"DefinitionList\"\u003e \u003cdiv class=\"DefinitionListEntry\"\u003e \u003cdiv class=\"Term\"\u003eAI\u003c/div\u003e \u003cdiv class=\"Description\"\u003e \u003cp\u003eArtificial Intelligence\u003c/p\u003e \u003c/div\u003e \u003c/div\u003e \u003cdiv class=\"DefinitionListEntry\"\u003e \u003cdiv class=\"Term\"\u003eAUC\u003c/div\u003e \u003cdiv class=\"Description\"\u003e \u003cp\u003eArea Under the ROC Curve\u003c/p\u003e \u003c/div\u003e \u003c/div\u003e \u003cdiv class=\"DefinitionListEntry\"\u003e \u003cdiv class=\"Term\"\u003eCBIS-DDSM\u003c/div\u003e \u003cdiv class=\"Description\"\u003e \u003cp\u003eCurated Breast Imaging Subset of DDSM\u003c/p\u003e \u003c/div\u003e \u003c/div\u003e \u003cdiv class=\"DefinitionListEntry\"\u003e \u003cdiv class=\"Term\"\u003eCSA\u003c/div\u003e \u003cdiv class=\"Description\"\u003e \u003cp\u003eCrow Search Algorithm\u003c/p\u003e \u003c/div\u003e \u003c/div\u003e \u003cdiv class=\"DefinitionListEntry\"\u003e \u003cdiv class=\"Term\"\u003eCV\u003c/div\u003e \u003cdiv class=\"Description\"\u003e \u003cp\u003eCross-Validation\u003c/p\u003e \u003c/div\u003e \u003c/div\u003e \u003cdiv class=\"DefinitionListEntry\"\u003e \u003cdiv class=\"Term\"\u003eDDSM\u003c/div\u003e \u003cdiv class=\"Description\"\u003e \u003cp\u003eDigital Database for Screening Mammography\u003c/p\u003e \u003c/div\u003e \u003c/div\u003e \u003cdiv class=\"DefinitionListEntry\"\u003e \u003cdiv class=\"Term\"\u003eGOA\u003c/div\u003e \u003cdiv class=\"Description\"\u003e \u003cp\u003eGrasshopper Optimization Algorithm\u003c/p\u003e \u003c/div\u003e \u003c/div\u003e \u003cdiv class=\"DefinitionListEntry\"\u003e \u003cdiv class=\"Term\"\u003eIBSI\u003c/div\u003e \u003cdiv class=\"Description\"\u003e \u003cp\u003eImage Biomarker Standardization Initiative\u003c/p\u003e \u003c/div\u003e \u003c/div\u003e \u003cdiv class=\"DefinitionListEntry\"\u003e \u003cdiv class=\"Term\"\u003eMLP\u003c/div\u003e \u003cdiv class=\"Description\"\u003e \u003cp\u003eMultilayer Perceptron\u003c/p\u003e \u003c/div\u003e \u003c/div\u003e \u003cdiv class=\"DefinitionListEntry\"\u003e \u003cdiv class=\"Term\"\u003eNPV\u003c/div\u003e \u003cdiv class=\"Description\"\u003e \u003cp\u003eNegative Predictive Value\u003c/p\u003e \u003c/div\u003e \u003c/div\u003e \u003cdiv class=\"DefinitionListEntry\"\u003e \u003cdiv class=\"Term\"\u003ePPV\u003c/div\u003e \u003cdiv class=\"Description\"\u003e \u003cp\u003ePositive Predictive Value\u003c/p\u003e \u003c/div\u003e \u003c/div\u003e \u003cdiv class=\"DefinitionListEntry\"\u003e \u003cdiv class=\"Term\"\u003eROC\u003c/div\u003e \u003cdiv class=\"Description\"\u003e \u003cp\u003eReceiver Operating Characteristic.\u003c/p\u003e \u003c/div\u003e \u003c/div\u003e \u003c/div\u003e"},{"header":"Declarations","content":"\u003cp\u003e \u003cstrong\u003eEthics approval and consent to participate\u003c/strong\u003e \u003cp\u003eNot applicable. This study used publicly available de-identified imaging datasets and synthetic data; no direct patient contact or intervention occurred.\u003c/p\u003e \u003c/p\u003e \u003cp\u003e \u003cstrong\u003eConsent for publication\u003c/strong\u003e \u003cp\u003eNot applicable.\u003c/p\u003e \u003c/p\u003e\u003cp\u003e \u003ch2\u003eCompeting interests\u003c/h2\u003e \u003cp\u003eThe author declares that there are no competing interests.\u003c/p\u003e\u003c/p\u003e\u003ch2\u003eFunding:\u003c/h2\u003e \u003cp\u003eNo specific funding was received for this work.\u003c/p\u003e\u003ch2\u003eAuthor Contribution\u003c/h2\u003e\u003cp\u003eThe author conceptualized the study, developed the methodology, curated the computational workflow, interpreted the results, and drafted the manuscript.\u003c/p\u003e\u003ch2\u003eAcknowledgement\u003c/h2\u003e\u003cp\u003eThe author would like to thank the Deanship of Scientific Research at Shaqra University for supporting this work. The author also acknowledges the providers of the publicly available CBIS-DDSM and VinDr-Mammo datasets used in this study.\u003c/p\u003e\u003ch2\u003eData Availability\u003c/h2\u003e\u003cp\u003eCBIS-DDSM and VinDr-Mammo are publicly available from their original repositories. The synthetic comparative experiment is generated from code included in the project repository. The code for feature extraction, feature selection, model training, and evaluation is available at https://github.com/bsalshreef/Hybrid-Metaheuristic-Mammography-AI.\u003c/p\u003e"},{"header":"References","content":"\u003col\u003e\u003cli\u003e\u003cspan\u003eChang YW et al. Artificial intelligence for breast cancer screening in mammography (AI-STREAM): preliminary analysis of a prospective multicenter cohort study. Nat Commun. 2025;[Epub ahead of print].\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eHernstr\u0026ouml;m V et al. Screening performance and characteristics of breast artificial intelligence in population-based mammography programmes. Lancet Digit Health. 2025;[Epub ahead of print].\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eL\u0026aring;ng K, et al. AI-supported breast cancer screening\u0026mdash;new results suggest even higher accuracy. Lund University report summarizing MASAI trial results; 2026.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eDembrower K et al. Nation-wide real-world implementation of AI for cancer detection in population-based mammography screening. Nat Med. 2025;[Epub ahead of print].\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eRadiological Society of North America. AI identifies interval breast cancer. RSNA News. 2025 Oct 27.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eGoh S, et al. Challenges in implementing artificial intelligence in breast cancer screening programs: qualitative insights from radiologists and administrators. J Med Internet Res. 2025;27:e62941.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eBreastcancer.org. Using AI (Artificial Intelligence) to Detect Breast Cancer. 2024 Mar 25. Available from: \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://www.breastcancer.org/\u003c/span\u003e\u003cspan address=\"https://www.breastcancer.org/\" targettype=\"URL\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eHadjievskyi N, et al. Robust radiomic feature selection in digital mammography: understanding the effect of imaging acquisition physics using phantom and clinical data analysis. Cancers (Basel). 2020;12(3):700.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eGillies RJ, Kinahan PE, Hricak H. Radiomics: images are more than pictures, they are data. Radiology. 2016;278(2):563\u0026ndash;77.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eAlsaadi T et al. MRI-based machine learning radiomics for prediction of breast cancer molecular subtypes. Diagn Interv Imaging. 2024;[Epub ahead of print].\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eLi X, et al. Advances in sampling, feature selection, and multi-omics integration for radiomics-based cancer prediction. Cancers (Basel). 2025;17(9):1610.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eBanu A, et al. Hybrid optimization algorithm based feature selection for mammogram images and detecting the breast mass using multilayer perceptron classifier. Comput Intell. 2022;38(3):888\u0026ndash;910.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eAhmed S, et al. Optimizing cancer classification: a metaheuristic-driven review of feature selection and deep neural network optimization (2012\u0026ndash;2025). Technol Cancer Res Treat. 2025;24:15330338251375817.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eKim H, et al. Artificial intelligence in breast ultrasound: a systematic review of clinical performance, bias, and implementation barriers. Front Oncol. 2025;15:1619364.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eJafari M, et al. Breast cancer prediction using hybrid machine learning and nature-inspired adaptive optimization algorithms. J Contemp Health Res. 2023;3(4):1\u0026ndash;10.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eErickson BJ, Korfiatis P, Akkus Z, Kline TL. Radiology artificial intelligence: radiological images and machine learning. Radiographics. 2019;39(7):1923\u0026ndash;39.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003ePark SH, Han K. Methodologic guide for evaluating clinical performance and effect of artificial intelligence technology for medical diagnosis and prediction. Radiology. 2018;286(3):800\u0026ndash;9.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eMongan J, Moy L, Kahn CE Jr. Checklist for artificial intelligence in medical imaging (CLAIM): a guide for authors and reviewers. Radiol Artif Intell. 2020;2(2):e200029.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eLambin P, et al. Radiomics: the bridge between medical imaging and personalized medicine. Nat Rev Clin Oncol. 2017;14(12):749\u0026ndash;62.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eYala A, et al. Toward robust mammography-based models for breast cancer risk. Sci Transl Med. 2019;11(575):eaax6883.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eTruhn D, et al. Radiomics in mammography: evaluation of feature stability and acquisition variability. Eur Radiol. 2021;31(5):3216\u0026ndash;27.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003ePesapane F, Rotili A, Penco S, Napolitano A, Houssami N, Sardanelli F. Artificial intelligence and mammography: where are we now? Eur J Radiol Exp. 2022;6(1):1\u0026ndash;18.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eChae EY, et al. Deep learning for breast cancer detection on digital mammograms: systematic review and meta-analysis. Eur J Radiol. 2023;161:110749.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eSechopoulos I, Mann RM. Stand-alone artificial intelligence for breast cancer detection in mammography: current evidence and real-world considerations. Nat Rev Clin Oncol. 2023;20(9):540\u0026ndash;56.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eDalmis MU, et al. Artificial intelligence for reducing workload in breast cancer screening with mammography: overview of current evidence. Eur J Radiol. 2024;165:111129.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eHoussami N et al. Will AI reshape breast cancer screening? ESMO Daily Reporter. 2025 May 11.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eLitjens G, et al. A survey on deep learning in medical image analysis. Med Image Anal. 2017;42:60\u0026ndash;88.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003evan Griethuysen JJM, et al. Computational radiomics system to decode the radiographic phenotype. Cancer Res. 2017;77(21):e104\u0026ndash;7.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eLee RS, Gimenez F, Hoogi A, et al. A curated mammography data set for use in computer-aided detection and diagnosis research. Sci Data. 2017;4:170177.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eClark K, Vendt B, Smith K, et al. The Cancer Imaging Archive (TCIA): maintaining and operating a public information repository. J Digit Imaging. 2013;26(6):1045\u0026ndash;57.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eNguyen T, et al. VinDr-Mammo: a large-scale benchmark dataset for computer-aided diagnosis in full-field digital mammography. Sci Data. 2023;10:262.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eAl-Helo S, et al. Narrative review of open-access mammography databases for artificial intelligence research. High Yield Med Rev. 2024;4(1):e24.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eLarson DB, et al. Ethics of using and sharing clinical imaging data for research and artificial intelligence development. Radiology. 2020;295(3):E131\u0026ndash;40.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eEthical design of. data-driven decision support tools for breast cancer screening. JMIR Cancer. 2025;e65566.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eFull Field Digital. Mammography Dataset from a Population-based Screening Program. Sci Data. 2025.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eZhou B, et al. Immunity-based Ebola optimization search algorithm for cancer classification using gene expression datasets. Sci Rep. 2022;12:18056.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eYoun I, et al. External validation of AI algorithms for breast cancer detection in digital mammography. Radiology. 2024;312(1):56\u0026ndash;67.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eRodriguez-Ruiz A, et al. Stand-alone artificial intelligence for breast cancer detection in screening mammography: comparison with 101 radiologists. J Natl Cancer Inst. 2019;111(9):916\u0026ndash;22.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eMcKinney SM, et al. International evaluation of an AI system for breast cancer screening. Nature. 2020;577(7788):89\u0026ndash;94.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eYala A, Lehman CD, Schuster T, Portnoi T, Barzilay R. A deep learning mammography-based model for improved breast cancer risk prediction. Radiology. 2019;292(1):60\u0026ndash;6.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eDuffy SW, et al. Randomized trials of mammography screening: updated overview of long-term mortality benefits. Radiology. 2020;296(2):221\u0026ndash;7.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eMann RM, Hooley RJ, Barr RG. Breast ultrasound in 2025 and beyond: impact of artificial intelligence and radiomics. AJR Am J Roentgenol. 2025;[Epub ahead of print].\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eTan M, Le Q. EfficientNet: rethinking model scaling for convolutional neural networks. Proc ICML. 2019;6105\u0026ndash;14.\u003c/span\u003e\u003c/li\u003e\u003c/ol\u003e"}],"fulltextSource":"","fullText":"","funders":[],"hasAdminPriorityOnWorkflow":false,"hasManuscriptDocX":true,"hasOptedInToPreprint":true,"hasPassedJournalQc":"","hasAnyPriority":true,"hideJournal":true,"highlight":"","institution":"","isAcceptedByJournal":false,"isAuthorSuppliedPdf":false,"isDeskRejected":"","isHiddenFromSearch":false,"isInQc":false,"isInWorkflow":false,"isPdf":false,"isPdfUpToDate":true,"isWithdrawnOrRetracted":false,"journal":{"display":true,"email":"[email protected]","identity":"researchsquare","isNatureJournal":false,"hasQc":true,"allowDirectSubmit":true,"externalIdentity":"","sideBox":"","snPcode":"","submissionUrl":"/submission","title":"Research Square","twitterHandle":"researchsquare","acdcEnabled":true,"dfaEnabled":false,"editorialSystem":"","reportingPortfolio":"","inReviewEnabled":false,"inReviewRevisionsEnabled":true},"keywords":"Breast cancer, artificial intelligence, digital mammography, radiomics, deep learning, feature selection, metaheuristic algorithms, decision support","lastPublishedDoi":"10.21203/rs.3.rs-9239930/v1","lastPublishedDoiUrl":"https://doi.org/10.21203/rs.3.rs-9239930/v1","license":{"name":"CC BY 4.0","url":"https://creativecommons.org/licenses/by/4.0/"},"manuscriptAbstract":"\u003ch2\u003eBackground\u003c/h2\u003e \u003cp\u003eArtificial intelligence (AI) shows promise for improving breast cancer detection in mammography, but generalizability across datasets and imaging conditions remains a major challenge. We developed a hybrid metaheuristic feature-selection framework that combines radiomics and deep learning features and evaluated it on a real pilot dataset and a controlled synthetic comparison.\u003c/p\u003e\u003ch2\u003eMethods\u003c/h2\u003e \u003cp\u003eA diagnostic model was developed using the public CBIS-DDSM dataset. The framework combined 2,051 IBSI-compliant radiomic features and 2,048-dimensional deep features from a pretrained EfficientNet-B5 model. A hybrid Grasshopper Optimization Algorithm and Crow Search Algorithm (GOA-CSA) was used to select an optimal feature subset for an MLP classifier. A controlled synthetic comparison (N\u0026thinsp;=\u0026thinsp;16, D\u0026thinsp;=\u0026thinsp;1114) compared an inventive multi-constraint fitness function against a legacy fitness under collapse-prone conditions.\u003c/p\u003e\u003ch2\u003eResults\u003c/h2\u003e \u003cp\u003eOn a CBIS-DDSM pilot subset (n\u0026thinsp;=\u0026thinsp;22, 5-fold cross-validation), the hybrid GOA-CSA model achieved an AUC of 0.858 while reducing the feature count by 95% to 102 features, compared with an all-features baseline AUC of 0.825. In the synthetic comparison, the inventive fitness achieved AUC 0.810 and sensitivity 0.571 versus 0.476 and 0.286 for the legacy fitness. The collapse-prevention mechanism was implemented but was not triggered in this synthetic run, as both models maintained sensitivity greater than zero.\u003c/p\u003e\u003ch2\u003eConclusions\u003c/h2\u003e \u003cp\u003eThe hybrid metaheuristic framework improved feature selection performance on both the real pilot and synthetic comparison. The synthetic experiment supports the value of the multi-constraint fitness design, but real-data validation of collapse prevention remains necessary.\u003c/p\u003e","manuscriptTitle":"Hybrid Metaheuristic Feature Selection for Enhanced Breast Cancer Detection in Digital Mammography: A Radiomics and Deep Learning Approach with Cross-Dataset Validation","msid":"","msnumber":"","nonDraftVersions":[{"code":1,"date":"2026-04-03 11:22:13","doi":"10.21203/rs.3.rs-9239930/v1","editorialEvents":[{"type":"communityComments","content":0}],"status":"published","journal":{"display":true,"email":"[email protected]","identity":"researchsquare","isNatureJournal":false,"hasQc":true,"allowDirectSubmit":true,"externalIdentity":"","sideBox":"","snPcode":"","submissionUrl":"/submission","title":"Research Square","twitterHandle":"researchsquare","acdcEnabled":true,"dfaEnabled":false,"editorialSystem":"","reportingPortfolio":"","inReviewEnabled":false,"inReviewRevisionsEnabled":true}}],"origin":"","ownerIdentity":"5908cd28-169e-4d19-af3b-34a2821ee1f0","owner":[],"postedDate":"April 3rd, 2026","published":true,"recentEditorialEvents":[],"rejectedJournal":[],"revision":"","amendment":"","status":"posted","subjectAreas":[],"tags":[],"updatedAt":"2026-04-21T01:23:41+00:00","versionOfRecord":[],"versionCreatedAt":"2026-04-03 11:22:13","video":"","vorDoi":"","vorDoiUrl":"","workflowStages":[]},"version":"v1","identity":"rs-9239930","journalConfig":"researchsquare"},"__N_SSP":true},"page":"/article/[identity]/[[...version]]","query":{"redirect":"/article/rs-9239930","identity":"rs-9239930","version":["v1"]},"buildId":"XKTyCvWXoU3ODBz1xrDgd","isFallback":false,"isExperimentalCompile":false,"dynamicIds":[84888],"gssp":true,"scriptLoader":[]}