18F-FDG PET/CT-Based Deep Radiomic Models for Enhancing Chemotherapy Response Prediction in Breast Cancer

18F-FDG PET/CT-Based Deep Radiomic Models for Enhancing Chemotherapy Response Prediction in Breast Cancer | 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 18F-FDG PET/CT-Based Deep Radiomic Models for Enhancing Chemotherapy Response Prediction in Breast Cancer Zirui Jiang, Joshua Low, Colin Huang, Yong Yue, Christopher Njeh, and 1 more This is a preprint; it has not been peer reviewed by a journal. https://doi.org/ 10.21203/rs.3.rs-6474899/v1 This work is licensed under a CC BY 4.0 License Status: Published Journal Publication published 11 Aug, 2025 Read the published version in Medical Oncology → Version 1 posted 10 You are reading this latest preprint version Abstract Background Enhancing the accuracy of tumor response predictions enables the development of tailored therapeutic strategies for patients with breast cancer. In this study, we developed deep-radiomic models to enhance the prediction of chemotherapy response after the first treatment cycle. Methods 18F- Fludeoxyglucose (FDG) PET/CT imaging data and clinical record from 60 breast cancer patients were retrospectively obtained from the Cancer Imaging Archive (QIN-BREAST). PET/CT scans were conducted at three distinct stages of treatment; prior to the initiation of chemotherapy (T1), following the first cycle of chemotherapy (T2), and after the full chemotherapy regimen (T3). The patient's primary gross tumor volume (GTV) was delineated on PET images using a 40% threshold of the maximum standardized uptake value (SUVmax). Radiomic features were extracted from the GTV based on the PET/CT images. In addition, a Squeeze-and-Excitation Network (SENet) deep learning model was employed to generate additional features from the PET/CT images for combined analysis. A XGBoost machine learning model was developed and compared with the conventional machine learning (ML) algorithm (random forest [RF], logistic regression [LR] and support vector machine [SVM]). The performance of each model was assessed using receiver operating characteristics area under the curve (ROC AUC) analysis, and prediction accuracy in a validation cohort. Model performance was evaluated through 5-fold cross-validation on the entire cohort, with data splits stratified by treatment response categories to ensure balanced representation. Results The AUC values for the machine learning models using only radiomic features were 0.85(XGBoost), 0.76 (RF), 0.80 (LR), and 0.59 (SVM), with XGBoost showing the best performance. After incorporating additional deep learning-derived features from SENet, the AUC values increased to 0.92, 0.88, 0.90, and 0.61, respectively, demonstrating significant improvements in predictive accuracy. Predictions were based on pre-treatment (T1) and post-first-cycle (T2) imaging data, enabling early assessment of chemotherapy response after the initial treatment cycle. Conclusion Integrating deep learning-derived features significantly enhanced the performance of predictive models for chemotherapy response in breast cancer patients. This study demonstrated the superior predictive capability of the XGBoost model, emphasizing its potential to optimize personalized therapeutic strategies by accurately identifying patients unlikely to respond to chemotherapy after the first treatment cycle. Figures Figure 1 Figure 2 Figure 3 Introduction Breast cancer is the most prevalent and consequential malignancies among women globally, posing a considerable challenge to public health [ 1 , 2 ]. Although there have been significant strides in treatment approaches, encompassing surgical intervention, radiation therapy, and chemotherapy, yet poor outcomes persist. The early prediction of patient response to treatment constitutes a critical yet unresolved issue [ 3 ]. The ability to identify patients unlikely to benefit from chemotherapy after the first treatment cycle has profound implications for oncologic management. It will enable the development of personalized treatment strategies, minimizes patient exposure to ineffective therapies, and optimizes the overall success of treatment regimens. Advanced imaging techniques, particularly the integration of Positron Emission Tomography/Computed Tomography (PET/CT) with machine learning models, represent a transformative development in oncologic management [ 4 ]. This integration harnesses the detailed imaging capabilities of PET/CT alongside the robust analytical power of machine learning to enhance the precision of predictive healthcare models. Recent studies have demonstrated that radiomic features extracted from imaging data can serve as valuable predictors of treatment response in various cancers [ 4 , 5 ]. The integration of machine learning (ML) into oncology, particularly in the realm of breast cancer, has significantly enhanced diagnostic precision and the prediction of treatment outcomes. Numerous studies have highlighted the effectiveness of ML models in processing complex, multidimensional datasets derived from clinical, imaging, and genetic sources, uncovering patterns and insights often overlooked by traditional analytical methods [ 4 , 6 ]. This advancement has revolutionized cancer diagnosis by enabling more nuanced interpretations of patient data, fostering personalized treatment strategies, optimizing therapeutic efficacy, and reducing adverse effects [ 4 , 6 – 8 ]. Furthermore, ML algorithms have played a pivotal role in risk stratification and prognostic evaluations, advancing personalized oncology by accurately predicting long-term patient outcomes and facilitating tailored clinical decision-making. Radiomics has emerged as a transformative approach in medical imaging, leveraging advanced machine learning techniques to extract high-dimensional features from medical images, such as those obtained from PET/CT scans [ 9 ]. These features, encompassing tumor shape, texture, and intensity, have been extensively utilized in studies aimed at developing predictive models for treatment outcomes in various cancers [ 9 , 10 ]. For instance, a systematic review and meta-analysis on esophageal cancer highlighted the efficacy of radiomics in predicting survival and treatment response, demonstrating the potential of combining clinical and radiomic features for improved prognostication [ 11 ]. Similarly, a study on hepatocellular carcinoma developed a radiomics-based model using multiparametric MRI, achieving enhanced predictive performance with a defined prediction score threshold [ 12 ]. These examples underscore the growing role of radiomics in integrating diverse imaging modalities to advance personalized treatment strategies and improve clinical outcomes. While radiomics has shown immense potential in predictive modeling, its application is not without limitations. One major challenge lies in the manual or semi-automated feature extraction process, which often relies on predefined radiomic features that may not fully capture the complexity and heterogeneity of tumors [ 13 ]. Additionally, radiomic models can be sensitive to variations in imaging protocols, reconstruction settings, and segmentation techniques, potentially affecting their reproducibility and generalizability across diverse clinical datasets [ 14 ]. To address these limitations, researchers have increasingly turned to deep learning-based approaches that extract deep features directly from medical images. To address these limitations, researchers have increasingly turned to deep learning-based approaches that extract deep features directly from medical images. Unlike traditional radiomics, which relies on handcrafted features, deep learning enables the automated extraction of high-dimensional, data-driven features that capture the complex and nuanced characteristics of tumors. Recent studies have demonstrated that integrating deep learning-extracted features with traditional radiomics methods can improve predictive accuracy and robustness in treatment response models. For example, Wang et al. (2024) developed a multimodality deep learning radiomics model that combined traditional and deep learning features from CT and MRI images to predict pathological complete response in esophageal squamous cell carcinoma, achieving an AUC of 0.868 [ 15 ]. Similarly, Wei et al. (2023) evaluated breast cancer axillary lymph node metastasis using deep learning radiomics of conventional ultrasound images, finding that the combination of deep learning and traditional radiomics features yielded an AUC of 0.92, outperforming models based solely on traditional radiomics [ 16 ]. These findings underscore the ability of deep learning-based approaches to comprehensively capture tumor heterogeneity, thus advancing the precision and reliability of predictive models for personalized oncology care. While the integration of deep learning-extracted features with traditional radiomics has shown success in other cancers, its application in breast cancer remains underexplored. Few studies have leveraged this combined approach to enhance predictive accuracy in breast cancer, leaving a gap in understanding its potential to improve treatment response predictions. This underscores the need for further research to adapt and validate these methodologies specifically for breast cancer cases. This study aimed to extract FDG PET/CT-based deep radiomic features and use it to develop ML model for predicting chemotherapy response in breast cancer patients. Additionally, this study compared the performance of eXtreme Gradient Boosting (XGBoost), with the traditional ML models such as SVM (Support Vector Machine), Logistic Regression (LR), and Random Forest (RF) to improve the prediction accuracy. Material and Methods Data Collection and Imaging pre-processing This retrospective study collected F-18 Fluorodeoxyglucose (FDG) PET/CT images and clinical notes from 60 breast cancer patients from the Cancer Imaging Archive (QIN-BREAST) [ 17 ]. Imaging data were collected at three pivotal treatment stages: prior to treatment initiation (T1), following the first chemotherapy cycle (T2), and after the completion of chemotherapy (T3). All patients had an average interval of 123 days, approximately four months, between the T2 and T3 imaging sessions. For the patients, the gross tumor volume (GTV) was delineated on pre-treatment PET images (T1) using a 40% threshold of the maximum standardized uptake value (SUVmax). Then, the GTV on the T1 images was co-registered with T2 and T3 PET/CT images to assess radiological response to treatment. Radiological responses were categorized based on reductions in PET avidity from baseline: a complete response (CR) was defined as a reduction of 75% or more, a partial response (PR) as a reduction between 40% and 74%, and no response (NR) as a reduction of less than 40%. These criteria align with the Positron Emission Tomography Response Criteria in Solid Tumors (PERCIST), which standardizes metabolic response assessments in oncology [ 18 ]. Implementing such standardized criteria enhances the consistency and accuracy of treatment response evaluations, facilitating better-informed clinical decisions. The imaging data employed in this study were sourced from the QIN-Breast dataset available through The Cancer Imaging Archive (TCIA). This dataset is publicly accessible and has been de-identified in compliance with the Health Insurance Portability and Accountability Act (HIPAA) regulations, ensuring the protection of patient privacy. As the data are anonymized and publicly available, the study was exempt from institutional review board (IRB) approval. Nonetheless, we adhered to all relevant guidelines and regulations governing the use of such data. In accordance with TCIA's data usage policies, we have cited the dataset appropriately and have not attempted to re-identify any individual participants. Software, Data Extraction, and Machine Learning Models: For this study, Python version 3.8 was used for code execution, model development, and debugging. Radiomic features were extracted from PET/CT images using Pyradiomics. Three machine learning models were implemented: XGBoost, RF, LR, and SVM, within the Python environment for a comprehensive assessment of results. All were conducted within the Python environment to facilitate a comprehensive assessment of the results. Radiomic Feature Extraction A comprehensive set of 314 radiomic features was extracted from each patient's primary tumor volume (Gross Tumor Volume, GTV) to facilitate the construction of machine-learning models. Specifically, 88 features were derived from the patients' CT imaging data, while the remaining 226 features originated from the PET imaging data. In addition to the radiomic features, a semi-quantitative analysis of PET parameters and clinical features, including the intervals between the three specified time points, was performed. To enhance the efficiency of the machine learning model computations and optimize the feature weights, all features were subjected to normalization. The performance of each machine learning model in predicting patient responses was assessed using the AUC analysis within a validation cohort. This approach ensured a robust evaluation of the predictive capabilities of the models. The machine learning model workflow is showed in Fig. 1. Deep Feature Extraction A deep learning model was also developed using the Segment-Anything Model (SAM), specifically employing the SAM_VIT_B_01EC64.pth to enhance the quality of the ROI in the imaging. Additionally, the SENet deep learning model (from SENet.PyTorch) was used to extract additional features from the GTV within the PET/CT images. Following feature extraction, the SelectKBest package was employed to identify the top 200 most significant features based on their predictive power. Deep learning-derived features were subsequently integrated with the original set of radiomic features to form a comprehensive dataset. Figure 1: the workflow of Deep-Radiomics Models Feature Selection and Validation To facilitate a rigorous evaluation of model performance, two distinct datasets were constructed: one comprising only the radiomic features and the other combining both the deep learning-derived features and the radiomic features. These datasets were used for training and evaluating multiple machine-learning models, allowing for a comparative analysis of the impact of incorporating deep learning-derived features on model accuracy. The predictive models were trained and validated using imaging data acquired at two critical timepoints: baseline (T1, pre-treatment) and post-first-cycle (T2). This design aligns with the clinical goal of early response prediction, allowing timely therapeutic adjustments before the second chemotherapy cycle. In this segment, patients were randomly allocated into training and testing sets, with a 5-fold cross-validation employed to ensure a reliable estimation of model performance. The training set was utilized for the training and refinement of models, while the testing set, isolated from the model fitting process, was used to evaluate the performance of the trained models. Results AUC values When using only radiomic features, the AUC values achieved by the RF, XGBoost, LR, and SVM models were 0.85(95%CI: 0.73–0.97), 0.76 (95%CI: 0.61–0.91), 0.80 (95%CI: 0.66–0.94), and 0.59 (95%CI: 0.41–0.77), respectively shown in Fig. 2 . Notably, only the SVM model failed to exceed the clinical threshold of 0.70, while the other models demonstrated strong predictive performance, with XGBoost achieving the highest AUC value of 0.84, highlighting its superior capability. After incorporating deep learning-extracted features with the radiomic features, the AUC values for XGBoost, LR, RF, and SVM increased to 0.92 (95%CI: 0.847-1.000), 0.90 (95%CI: 0.818–0.982), 0.88 (95%CI: 0.791–0.969), and 0.61 (95%CI: 0.467–0.753), respectively, as shown in Fig. 2 . While the performance of the SVM model remained suboptimal, the AUC values for the other models demonstrated significant improvements, underscoring the value of integrating deep learning-derived features with radiomic features. Statistical significance of AUC differences between models was assessed using DeLong’s test. After Bonferroni correction for multiple comparisons (α = 0.05/6 pairwise comparisons), the improvement in XGBoost’s AUC from 0.85 (radiomics-only) to 0.92 (radiomics + deep features) was statistically significant (p = 0.003). Similarly, LR and RF showed significant improvements (p = 0.012 and p = 0.021, respectively), while SVM’s performance remained unchanged (p = 0.42). Table 1 Performance comparison of radiomics-only and combined models using ROC analysis Radiomics Only (AUC) Combined (AUC) DeLong’s test (p-value) XGBoost 0.85 0.92 0.003* Logistic Regression 0.80 0.90 0.012* Random Forest 0.76 0.87 0.021* SVM 0.59 0.61 0.42 *Note: AUC = Area Under the Curve; * indicates statistical significance (p < 0.05). Decision Curve Analysis Decision curve analysis revealed distinct net benefit patterns across threshold probabilities for the three models (Figure X). The XGBoost model achieved the highest net benefit at low thresholds (0.05–0.30), maintaining stable performance without significant fluctuations. In contrast, logistic regression dominated the mid-threshold range (0.30–0.55), with a peak net benefit of 0.42 at 0.40, but exhibited declining utility beyond 0.55, reaching negative net benefits at thresholds > 0.70. Random forest demonstrated marked variability, with transient peaks at 0.25 (net benefit: 0.38) and 0.50 (0.35), but sharp declines in adjacent ranges (e.g., 0.30–0.45 and 0.55–0.65). As baseline references, the "Treat All" and "Treat None" strategies were outperformed by XGBoost and LR in their respective effective ranges, further demonstrating the potential clinical utility of these models in optimizing treatment decisions. Discussion This study highlighted the effectiveness of integrating PET/CT imaging data with machine learning models to predict early chemotherapy responses in breast cancer patients. Among the models evaluated, XGBoost demonstrated the highest predictive capability, achieving an Area AUC value of 0.85. This performance exceeded that of RF (AUC: 0.76), LR (AUC: 0.80), and SVM (AUC: 0.59). The superior performance of the XGBoost model underscores its potential as a powerful tool for early chemotherapy response prediction, enabling more precise and personalized treatment strategies for breast cancer patients. In this context, XGBoost emerges as a robust solution capable of addressing these limitations. XGBoost is designed to handle large-scale datasets with superior computational efficiency and enhanced resistance to overfitting [ 19 , 20 ]. This algorithm applies gradient boosting principles to iteratively refine prediction accuracy, making it adept at managing complex interactions within high-dimensional data. Therefore, applying XGBoost in predicting breast cancer response should be a key area of focus in current research. Additionally, the application of deep learning models such as the Squeeze-and-Excitation Network (SENet) further enhances the ability to capture significant imaging features from PET/CT data, providing more robust inputs for predictive modeling [ 21 ]. By incorporating deep learning-derived features extracted from PET/CT images using the SENet model, the predictive performance of all machine learning models markedly improved. The AUC values for XGBoost, LR, RF, and SVM increased to 0.92, 0.90, 0.88, and 0.61, respectively. These results underscore the significant value of integrating deep learning-extracted features with traditional radiomic features, enhancing the precision and reliability of treatment planning. While the SVM model continued to underperform compared to other models, the substantial improvement in the XGBoost, LR, and RF models highlights the potential of deep learning-derived features to augment the predictive power of conventional radiomic analysis. The statistically validated superiority of XGBoost underscores its capability to leverage both radiomic and deep learning-derived features, whereas SVM’s limitations in handling nonlinear interactions may explain its suboptimal performance. The superior performance of the XGBoost model can be largely attributed to its sophisticated algorithmic design, which effectively handles complex interactions among features and exhibits robustness in managing high-dimensional datasets [ 22 ]. This model’s predictive accuracy is significantly bolstered by its capacity to incorporate and analyze a diverse array of radiomic features. Specifically, the Interquartile Range of the GTV at the pre-treatment stage (T1) and the difference in Mean Standardized Uptake Value (SUVmean) between the pre-treatment stage (T1) and the post-first chemotherapy cycle (T2) emerged as critical predictors. These features likely capture essential variations in tumor metabolism and morphological changes induced by chemotherapy, which are pivotal for assessing treatment efficacy. The detailed and dynamic nature of these predictors contributes to the enhanced reliability and accuracy of the XGBoost model. Furthermore, the model’s ability to integrate semi-quantitative PET parameters and clinical features underscores its comprehensive approach to predicting chemotherapy response, thereby reinforcing its potential as a valuable tool in the realm of personalized cancer treatment. A potential critique of our methodology lies in its deviation from the conventional use of DL models, which are typically employed to predict outcomes directly without explicit feature extraction. In contrast, our approach utilized DL primarily for feature extraction, effectively reframing it into a handcrafted feature model. While this approach may seem unconventional, it enables a more interpretable integration of deep learning-derived features with machine learning models, allowing for a clearer understanding of which features are most predictive of chemotherapy response. Importantly, we speculate that automating the entire DL process—without a separate feature extraction step—would not result in a degradation of model performance. In fact, we expect automation to further enhance predictive accuracy and efficiency, particularly as the model continues to evolve with larger datasets and improved optimization techniques. By combining DL-derived features with traditional radiomic analysis, our methodology offers a flexible and scalable framework, capable of adapting to future advancements in medical imaging and DL technologies. Notably, for DCA curve, although both the XGBoost and Random Forest models outperformed the "Treat All" and "Treat None" strategies across a substantial portion of the threshold probability range, demonstrating their potential value in clinical decision-making. However, certain limitations of the DCA curves warrant attention. The curves exhibited noticeable fluctuations, likely due to the limited sample size of only 15 patients, which may have introduced instability and amplified the impact of data noise on model performance. Additionally, at lower threshold probabilities (e.g., 0.5), all curves closely aligned with the "Treat All" and "Treat None" baselines, indicating that the models offered limited decision-making value under extreme conditions. To address these shortcomings in future studies, we aim to expand the dataset to reduce variability, enhance the stability of the DCA curves, and achieve smoother, more reliable performance metrics. Our findings are consistent with previous research that underscores the potential of machine learning models in predicting cancer treatment outcomes. Several studies have illustrated how the integration of imaging data with machine learning algorithms can enhance predictive accuracy, thereby improving treatment planning and patient management. For instance, studies by Cai et al. (2023) and Zhang et al. (2018) have demonstrated that radiomic features extracted from imaging data can serve as valuable predictors of treatment response in various cancers [ 7 , 23 ]. The strength of our approach lies in the comprehensive integration of PET/CT imaging data with machine learning models, which allows for a more nuanced and detailed analysis of tumor characteristics and treatment responses. This methodology leverages the strengths of both imaging and computational analysis, facilitating a deeper understanding of the factors that influence chemotherapy efficacy. However, several potential limitations must be acknowledged. The retrospective design of our study may introduce inherent biases, as the data were collected and analyzed after the fact, which can affect the interpretation of causality and outcome relationships. Additionally, a large sample size would enhance generalizability of our findings, providing a more robust validation of the model's predictive power. The inclusion of a diverse patient population and multi-institutional data could further enhance the reliability and applicability of the results. Future research should focus on prospective studies with larger sample sizes to validate our findings and address these limitations. Moreover, incorporating additional predictive variables, such as genomic data and more comprehensive clinical features, could potentially enhance the models' accuracy and applicability, paving the way for more personalized and effective cancer treatment strategies. The ability to predict chemotherapy responses early in the treatment process holds substantial clinical implications. Early identification of patients unlikely to respond to standard chemotherapy regimens enables clinicians to adjust treatment plans promptly, potentially opting for alternative therapies such as radiation or hormone therapy. This personalized approach optimizes treatment efficacy and minimizes unnecessary exposure to ineffective treatments, thereby improving patient outcomes and reducing healthcare costs [ 24 , 25 ]. Early prediction allows for a more tailored therapeutic strategy, ensuring that each patient receives the most appropriate and effective treatment based on their specific tumor characteristics and predicted response. Future research should focus on prospective studies to validate the predictive models developed in this study. Additionally, incorporating more predictive variables, such as genomic data and additional radiomic features, could enhance the models' accuracy and applicability. Testing these models in real-world clinical settings will also be crucial to assess their practical utility and refine them based on clinical feedback. Conclusion In conclusion, our study demonstrated that integrating PET/CT imaging data with machine learning models, including XGBoost, SVM, LR, and RF model, effectively predicts early chemotherapy responses in breast cancer patients. The XGBoost model showed superior predictive capability, facilitating personalized treatment strategies. Early identification of non-responders allows for timely interventional therapy, enhancing treatment efficacy and avoiding unnecessary continuation of ineffective treatments. This approach underscores the potential of advanced machine learning models in improving patient outcomes and optimizing cancer treatment management. Declarations Funding OO– Supported by Purdue University Research Funding. Human Ethics and Consent to Participate declarations The imaging data employed in this study were sourced from the QIN-Breast dataset available through The Cancer Imaging Archive (TCIA). This dataset is publicly accessible and has been de-identified in compliance with the Health Insurance Portability and Accountability Act (HIPAA) regulations, ensuring the protection of patient privacy. As the data are anonymized and publicly available, the study was exempt from institutional review board (IRB) approval. Nonetheless, we adhered to all relevant guidelines and regulations governing the use of such data. In accordance with TCIA's data usage policies, we have cited the dataset appropriately and have not attempted to re-identify any individual participants. Competing interests The authors declare that they have no competing interests. Author Contribution Z.J. conceptualized the study, designed the methodology, performed radiomic analysis, developed machine learning models, and wrote the original manuscript. J.L. contributed to data collection, preprocessing, and preliminary statistical validation. C.H., Y.Y., and C.N. provided technical resources and reviewed the manuscript. O.O. supervised the project, provided critical revisions to the manuscript, and approved the final version. All authors read and approved the final manuscript. References Sung H et al. Global cancer statistics 2020: GLOBOCAN estimates of incidence and mortality worldwide for 36 cancers in 185 countries. 2021. 71(3): pp. 209–49. Łukasiewicz S et al. Breast cancer—epidemiology, risk factors, classification, prognostic markers, and current treatment strategies—an updated review. 2021. 13(17): p. 4287. Fang S et al. A real-world clinicopathological model for predicting pathological complete response to neoadjuvant chemotherapy in breast cancer. 2024. 14: p. 1323226. Hadebe B et al. The role of PET/CT in breast cancer. 2023. 13(4): p. 597. Li T et al. Applications of FAPI PET/CT in the diagnosis and treatment of breast and the most common gynecologic malignancies: a literature review. 2024. 14: p. 1358070. Iqbal MJ, et al. Clin Appl Artif Intell Mach Learn cancer diagnosis: Look into future. 2021;21(1):270. Zhang B, Shi H. and H.J.J.o.m.h. Wang, Machine learning and AI in cancer prognosis, prediction, and treatment selection: a critical approach. 2023: pp. 1779–1791. Rasool A et al. Improved machine learning-based predictive models for breast cancer diagnosis. 2022. 19(6): p. 3211. Zhang Y et al. PET radiomics in lung cancer: advances and translational challenges. 2024. 11(1): p. 81. Zhou Y et al. Integrating 18F-FDG PET/CT radiomics and body composition for enhanced prognostic assessment in patients with esophageal cancer. 2024. 24(1): p. 1402. Menon N et al. Performance of radiomics-based artificial intelligence systems in the diagnosis and prediction of treatment response and survival in esophageal cancer: a systematic review and meta-analysis of diagnostic accuracy. 2023. 36(6): p. doad034. Sim Y et al. Multiparametric MRI–based radiomics model for predicting human papillomavirus status in oropharyngeal squamous cell carcinoma: optimization using oversampling and machine learning techniques. 2024. 34(5): pp. 3102–12. Lohmann P et al. Radiomics in radiation oncology—basics, methods, and limitations. 2020. 196(10): pp. 848–855. Mali SA et al. Making radiomics more reproducible across scanner and imaging protocol variations: a review of harmonization methods. 2021. 11(9): p. 842. Liu Y et al. Multimodality deep learning radiomics predicts pathological response after neoadjuvant chemoradiotherapy for esophageal squamous cell carcinoma. 2024. 15(1): p. 277. Wei W et al. Deep learning radiomics for prediction of axillary lymph node metastasis in patients with clinical stage T1–2 breast cancer. 2023. 13(8): p. 4995. Li X et al. Data QIN-breast 2016. 5. Wahl RL et al. From RECIST to PERCIST: evolving considerations for PET response criteria in solid tumors. 2009. 50(Suppl 1): p. 122S-150S. Zhang X et al. Predicting missing values in medical data via XGBoost regression. 2020. 4: pp. 383–394. Xu Y et al. Predicting patients' satisfaction with doctors in online medical communities: An approach based on XGBoost algorithm. 2022. 34(4): pp. 1–17. Hu J, Shen L, Sun G. Squeeze-and-excitation networks . in Proceedings of the IEEE conference on computer vision and pattern recognition . 2018. Ma B, et al. XGBLC: improved survival prediction model based XGBoost. 2022;38(2):410–8. Cai J et al. A radiomics model for predicting the response to bevacizumab in brain necrosis after radiotherapy. 2020. 26(20): pp. 5438–47. Romeo V et al. Assessment and prediction of response to neoadjuvant chemotherapy in breast cancer: A comparison of imaging modalities and future perspectives. 2021. 13(14): p. 3521. Liang X, et al. Early prediction of pathological complete response to neoadjuvant chemotherapy combining DCE-MRI and apparent diffusion coefficient values. Breast Cancer. 2022;22(1):1250. Additional Declarations No competing interests reported. Cite Share Download PDF Status: Published Journal Publication published 11 Aug, 2025 Read the published version in Medical Oncology → Version 1 posted Editorial decision: Revision requested 12 Jul, 2025 Reviews received at journal 30 Jun, 2025 Reviewers agreed at journal 21 Jun, 2025 Reviewers agreed at journal 20 Jun, 2025 Reviews received at journal 20 May, 2025 Reviewers agreed at journal 12 May, 2025 Reviewers invited by journal 11 May, 2025 Editor assigned by journal 18 Apr, 2025 Submission checks completed at journal 18 Apr, 2025 First submitted to journal 17 Apr, 2025 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-6474899","acceptedTermsAndConditions":true,"allowDirectSubmit":false,"archivedVersions":[],"articleType":"Research Article","associatedPublications":[],"authors":[{"id":455399786,"identity":"ab27de09-08eb-439a-ad19-307994d616ee","order_by":0,"name":"Zirui Jiang","email":"","orcid":"","institution":"Purdue University","correspondingAuthor":false,"prefix":"","firstName":"Zirui","middleName":"","lastName":"Jiang","suffix":""},{"id":455399787,"identity":"412eae90-7223-4f72-b1ae-80b5384c30e5","order_by":1,"name":"Joshua Low","email":"","orcid":"","institution":"Purdue University","correspondingAuthor":false,"prefix":"","firstName":"Joshua","middleName":"","lastName":"Low","suffix":""},{"id":455399788,"identity":"aaa6ddf7-ea6c-4fa6-bd75-1d44865f4585","order_by":2,"name":"Colin Huang","email":"","orcid":"","institution":"Indiana University School of Medicine","correspondingAuthor":false,"prefix":"","firstName":"Colin","middleName":"","lastName":"Huang","suffix":""},{"id":455399789,"identity":"2d166acc-b768-4347-9997-675239a54322","order_by":3,"name":"Yong Yue","email":"","orcid":"","institution":"Indiana University School of Medicine","correspondingAuthor":false,"prefix":"","firstName":"Yong","middleName":"","lastName":"Yue","suffix":""},{"id":455399790,"identity":"ab4bf06c-af9d-4091-bd2c-c35ab9b97a09","order_by":4,"name":"Christopher Njeh","email":"","orcid":"","institution":"Indiana University School of Medicine","correspondingAuthor":false,"prefix":"","firstName":"Christopher","middleName":"","lastName":"Njeh","suffix":""},{"id":455399791,"identity":"c1a5f1ea-d974-4a2e-9906-a30e931befba","order_by":5,"name":"Oluwaseyi Oderinde","email":"data:image/png;base64,iVBORw0KGgoAAAANSUhEUgAAAZAAAAAyAQMAAABI0h/eAAAABlBMVEX///8AAABVwtN+AAAACXBIWXMAAA7EAAAOxAGVKw4bAAAAyElEQVRIiWNgGAWjYHACAxBi7AcxecACCURqmdlAmhYGBsYNB4jVws/evPFxRcEd2c23e8w+vG2zAYrkGODVItlzrNjwjMEz4213zhjPnNuWBhR5g1+LwY0cM8kGg8OJ227kGDPzth0GiRDUYv4TpGXzDLCW/wz2RGgxYwRp2SAB1nKAwUCCCL+AHGY8486xYsY555J5JM48K8CrBRRiHxv+HJbtn928meFNmZ0cf3vyBrxaEEACQvEQqRxJyygYBaNgFIwCDAAA7gBIGoF2nwAAAAAASUVORK5CYII=","orcid":"","institution":"Purdue University","correspondingAuthor":true,"prefix":"","firstName":"Oluwaseyi","middleName":"","lastName":"Oderinde","suffix":""}],"badges":[],"createdAt":"2025-04-18 00:23:06","currentVersionCode":1,"declarations":"","doi":"10.21203/rs.3.rs-6474899/v1","doiUrl":"https://doi.org/10.21203/rs.3.rs-6474899/v1","draftVersion":[],"editorialEvents":[{"content":"https://doi.org/10.1007/s12032-025-02982-0","type":"published","date":"2025-08-11T15:57:29+00:00"}],"editorialNote":"","failedWorkflow":false,"files":[{"id":82795428,"identity":"a21562cb-1102-427b-92c9-5febe7f073b8","added_by":"auto","created_at":"2025-05-15 10:35:22","extension":"png","order_by":1,"title":"Figure 1","display":"","copyAsset":false,"role":"figure","size":115441,"visible":true,"origin":"","legend":"\u003cp\u003e\u003cem\u003ethe workflow of Deep-Radiomics Models\u003c/em\u003e\u003c/p\u003e","description":"","filename":"floatimage1.png","url":"https://assets-eu.researchsquare.com/files/rs-6474899/v1/236bf4d00cbfec506b61cffc.png"},{"id":82798540,"identity":"a1754e3d-bb30-40c6-b373-2b87ee93ec57","added_by":"auto","created_at":"2025-05-15 10:51:22","extension":"png","order_by":2,"title":"Figure 2","display":"","copyAsset":false,"role":"figure","size":74484,"visible":true,"origin":"","legend":"\u003cp\u003e\u003cem\u003eComparison of Radiomics-Only and Combined Model Performance Using ROC Curves\u003c/em\u003e\u003c/p\u003e","description":"","filename":"floatimage2.png","url":"https://assets-eu.researchsquare.com/files/rs-6474899/v1/105ef33c8a1ab599dfb91163.png"},{"id":82795424,"identity":"35cd7f7a-7ace-4f1b-adf5-d348494ae768","added_by":"auto","created_at":"2025-05-15 10:35:22","extension":"png","order_by":3,"title":"Figure 3","display":"","copyAsset":false,"role":"figure","size":119539,"visible":true,"origin":"","legend":"\u003cp\u003e\u003cem\u003eFigure 6: Decision Curve Analysis (XGBoost vs Logistic Regression vs Random Forest)\u003c/em\u003e\u003c/p\u003e\n\u003cp\u003eAs baseline references, the \"Treat All\" and \"Treat None\" strategies were outperformed by XGBoost and LR in their respective effective ranges, further demonstrating the potential clinical utility of these models in optimizing treatment decisions.\u003c/p\u003e","description":"","filename":"floatimage3.png","url":"https://assets-eu.researchsquare.com/files/rs-6474899/v1/f8a749b94422a8d7dc53663f.png"},{"id":89310542,"identity":"96c2c738-216b-4358-b3ee-7084fca863ee","added_by":"auto","created_at":"2025-08-18 16:07:33","extension":"pdf","order_by":0,"title":"","display":"","copyAsset":false,"role":"manuscript-pdf","size":821177,"visible":true,"origin":"","legend":"","description":"","filename":"manuscript.pdf","url":"https://assets-eu.researchsquare.com/files/rs-6474899/v1/5514b406-70e0-42f9-a924-76de1734a450.pdf"}],"financialInterests":"No competing interests reported.","formattedTitle":"18F-FDG PET/CT-Based Deep Radiomic Models for Enhancing Chemotherapy Response Prediction in Breast Cancer","fulltext":[{"header":"Introduction","content":"\u003cp\u003eBreast cancer is the most prevalent and consequential malignancies among women globally, posing a considerable challenge to public health [\u003cspan citationid=\"CR1\" class=\"CitationRef\"\u003e1\u003c/span\u003e, \u003cspan citationid=\"CR2\" class=\"CitationRef\"\u003e2\u003c/span\u003e]. Although there have been significant strides in treatment approaches, encompassing surgical intervention, radiation therapy, and chemotherapy, yet poor outcomes persist. The early prediction of patient response to treatment constitutes a critical yet unresolved issue [\u003cspan citationid=\"CR3\" class=\"CitationRef\"\u003e3\u003c/span\u003e]. The ability to identify patients unlikely to benefit from chemotherapy after the first treatment cycle has profound implications for oncologic management. It will enable the development of personalized treatment strategies, minimizes patient exposure to ineffective therapies, and optimizes the overall success of treatment regimens.\u003c/p\u003e \u003cp\u003eAdvanced imaging techniques, particularly the integration of Positron Emission Tomography/Computed Tomography (PET/CT) with machine learning models, represent a transformative development in oncologic management [\u003cspan citationid=\"CR4\" class=\"CitationRef\"\u003e4\u003c/span\u003e]. This integration harnesses the detailed imaging capabilities of PET/CT alongside the robust analytical power of machine learning to enhance the precision of predictive healthcare models. Recent studies have demonstrated that radiomic features extracted from imaging data can serve as valuable predictors of treatment response in various cancers [\u003cspan citationid=\"CR4\" class=\"CitationRef\"\u003e4\u003c/span\u003e, \u003cspan citationid=\"CR5\" class=\"CitationRef\"\u003e5\u003c/span\u003e].\u003c/p\u003e \u003cp\u003eThe integration of machine learning (ML) into oncology, particularly in the realm of breast cancer, has significantly enhanced diagnostic precision and the prediction of treatment outcomes. Numerous studies have highlighted the effectiveness of ML models in processing complex, multidimensional datasets derived from clinical, imaging, and genetic sources, uncovering patterns and insights often overlooked by traditional analytical methods [\u003cspan citationid=\"CR4\" class=\"CitationRef\"\u003e4\u003c/span\u003e, \u003cspan citationid=\"CR6\" class=\"CitationRef\"\u003e6\u003c/span\u003e]. This advancement has revolutionized cancer diagnosis by enabling more nuanced interpretations of patient data, fostering personalized treatment strategies, optimizing therapeutic efficacy, and reducing adverse effects [\u003cspan citationid=\"CR4\" class=\"CitationRef\"\u003e4\u003c/span\u003e, \u003cspan additionalcitationids=\"CR7\" citationid=\"CR6\" class=\"CitationRef\"\u003e6\u003c/span\u003e\u0026ndash;\u003cspan citationid=\"CR8\" class=\"CitationRef\"\u003e8\u003c/span\u003e]. Furthermore, ML algorithms have played a pivotal role in risk stratification and prognostic evaluations, advancing personalized oncology by accurately predicting long-term patient outcomes and facilitating tailored clinical decision-making.\u003c/p\u003e \u003cp\u003eRadiomics has emerged as a transformative approach in medical imaging, leveraging advanced machine learning techniques to extract high-dimensional features from medical images, such as those obtained from PET/CT scans [\u003cspan citationid=\"CR9\" class=\"CitationRef\"\u003e9\u003c/span\u003e]. These features, encompassing tumor shape, texture, and intensity, have been extensively utilized in studies aimed at developing predictive models for treatment outcomes in various cancers [\u003cspan citationid=\"CR9\" class=\"CitationRef\"\u003e9\u003c/span\u003e, \u003cspan citationid=\"CR10\" class=\"CitationRef\"\u003e10\u003c/span\u003e]. For instance, a systematic review and meta-analysis on esophageal cancer highlighted the efficacy of radiomics in predicting survival and treatment response, demonstrating the potential of combining clinical and radiomic features for improved prognostication [\u003cspan citationid=\"CR11\" class=\"CitationRef\"\u003e11\u003c/span\u003e]. Similarly, a study on hepatocellular carcinoma developed a radiomics-based model using multiparametric MRI, achieving enhanced predictive performance with a defined prediction score threshold [\u003cspan citationid=\"CR12\" class=\"CitationRef\"\u003e12\u003c/span\u003e]. These examples underscore the growing role of radiomics in integrating diverse imaging modalities to advance personalized treatment strategies and improve clinical outcomes.\u003c/p\u003e \u003cp\u003eWhile radiomics has shown immense potential in predictive modeling, its application is not without limitations. One major challenge lies in the manual or semi-automated feature extraction process, which often relies on predefined radiomic features that may not fully capture the complexity and heterogeneity of tumors [\u003cspan citationid=\"CR13\" class=\"CitationRef\"\u003e13\u003c/span\u003e]. Additionally, radiomic models can be sensitive to variations in imaging protocols, reconstruction settings, and segmentation techniques, potentially affecting their reproducibility and generalizability across diverse clinical datasets [\u003cspan citationid=\"CR14\" class=\"CitationRef\"\u003e14\u003c/span\u003e]. To address these limitations, researchers have increasingly turned to deep learning-based approaches that extract deep features directly from medical images.\u003c/p\u003e \u003cp\u003eTo address these limitations, researchers have increasingly turned to deep learning-based approaches that extract deep features directly from medical images. Unlike traditional radiomics, which relies on handcrafted features, deep learning enables the automated extraction of high-dimensional, data-driven features that capture the complex and nuanced characteristics of tumors. Recent studies have demonstrated that integrating deep learning-extracted features with traditional radiomics methods can improve predictive accuracy and robustness in treatment response models. For example, Wang et al. (2024) developed a multimodality deep learning radiomics model that combined traditional and deep learning features from CT and MRI images to predict pathological complete response in esophageal squamous cell carcinoma, achieving an AUC of 0.868 [\u003cspan citationid=\"CR15\" class=\"CitationRef\"\u003e15\u003c/span\u003e]. Similarly, Wei et al. (2023) evaluated breast cancer axillary lymph node metastasis using deep learning radiomics of conventional ultrasound images, finding that the combination of deep learning and traditional radiomics features yielded an AUC of 0.92, outperforming models based solely on traditional radiomics [\u003cspan citationid=\"CR16\" class=\"CitationRef\"\u003e16\u003c/span\u003e]. These findings underscore the ability of deep learning-based approaches to comprehensively capture tumor heterogeneity, thus advancing the precision and reliability of predictive models for personalized oncology care.\u003c/p\u003e \u003cp\u003eWhile the integration of deep learning-extracted features with traditional radiomics has shown success in other cancers, its application in breast cancer remains underexplored. Few studies have leveraged this combined approach to enhance predictive accuracy in breast cancer, leaving a gap in understanding its potential to improve treatment response predictions. This underscores the need for further research to adapt and validate these methodologies specifically for breast cancer cases. This study aimed to extract FDG PET/CT-based deep radiomic features and use it to develop ML model for predicting chemotherapy response in breast cancer patients. Additionally, this study compared the performance of eXtreme Gradient Boosting (XGBoost), with the traditional ML models such as SVM (Support Vector Machine), Logistic Regression (LR), and Random Forest (RF) to improve the prediction accuracy.\u003c/p\u003e"},{"header":"Material and Methods","content":"\u003cdiv id=\"Sec3\" class=\"Section2\"\u003e \u003ch2\u003eData Collection and Imaging pre-processing\u003c/h2\u003e \u003cp\u003eThis retrospective study collected F-18 Fluorodeoxyglucose (FDG) PET/CT images and clinical notes from 60 breast cancer patients from the Cancer Imaging Archive (QIN-BREAST) [\u003cspan citationid=\"CR17\" class=\"CitationRef\"\u003e17\u003c/span\u003e]. Imaging data were collected at three pivotal treatment stages: prior to treatment initiation (T1), following the first chemotherapy cycle (T2), and after the completion of chemotherapy (T3). All patients had an average interval of 123 days, approximately four months, between the T2 and T3 imaging sessions. For the patients, the gross tumor volume (GTV) was delineated on pre-treatment PET images (T1) using a 40% threshold of the maximum standardized uptake value (SUVmax). Then, the GTV on the T1 images was co-registered with T2 and T3 PET/CT images to assess radiological response to treatment. Radiological responses were categorized based on reductions in PET avidity from baseline: a complete response (CR) was defined as a reduction of 75% or more, a partial response (PR) as a reduction between 40% and 74%, and no response (NR) as a reduction of less than 40%. These criteria align with the Positron Emission Tomography Response Criteria in Solid Tumors (PERCIST), which standardizes metabolic response assessments in oncology [\u003cspan citationid=\"CR18\" class=\"CitationRef\"\u003e18\u003c/span\u003e]. Implementing such standardized criteria enhances the consistency and accuracy of treatment response evaluations, facilitating better-informed clinical decisions. The imaging data employed in this study were sourced from the QIN-Breast dataset available through The Cancer Imaging Archive (TCIA). This dataset is publicly accessible and has been de-identified in compliance with the Health Insurance Portability and Accountability Act (HIPAA) regulations, ensuring the protection of patient privacy. As the data are anonymized and publicly available, the study was exempt from institutional review board (IRB) approval. Nonetheless, we adhered to all relevant guidelines and regulations governing the use of such data. In accordance with TCIA's data usage policies, we have cited the dataset appropriately and have not attempted to re-identify any individual participants.\u003c/p\u003e \u003c/div\u003e\n\u003ch3\u003eSoftware, Data Extraction, and Machine Learning Models:\u003c/h3\u003e\n\u003cp\u003eFor this study, Python version 3.8 was used for code execution, model development, and debugging. Radiomic features were extracted from PET/CT images using Pyradiomics. Three machine learning models were implemented: XGBoost, RF, LR, and SVM, within the Python environment for a comprehensive assessment of results. All were conducted within the Python environment to facilitate a comprehensive assessment of the results.\u003c/p\u003e\n\u003ch3\u003eRadiomic Feature Extraction\u003c/h3\u003e\n\u003cp\u003eA comprehensive set of 314 radiomic features was extracted from each patient's primary tumor volume (Gross Tumor Volume, GTV) to facilitate the construction of machine-learning models. Specifically, 88 features were derived from the patients' CT imaging data, while the remaining 226 features originated from the PET imaging data. In addition to the radiomic features, a semi-quantitative analysis of PET parameters and clinical features, including the intervals between the three specified time points, was performed.\u003c/p\u003e \u003cp\u003eTo enhance the efficiency of the machine learning model computations and optimize the feature weights, all features were subjected to normalization. The performance of each machine learning model in predicting patient responses was assessed using the AUC analysis within a validation cohort. This approach ensured a robust evaluation of the predictive capabilities of the models. The machine learning model workflow is showed in Fig.\u0026nbsp;1.\u003c/p\u003e\n\u003ch3\u003eDeep Feature Extraction\u003c/h3\u003e\n\u003cp\u003eA deep learning model was also developed using the Segment-Anything Model (SAM), specifically employing the SAM_VIT_B_01EC64.pth to enhance the quality of the ROI in the imaging. Additionally, the SENet deep learning model (from SENet.PyTorch) was used to extract additional features from the GTV within the PET/CT images. Following feature extraction, the SelectKBest package was employed to identify the top 200 most significant features based on their predictive power. Deep learning-derived features were subsequently integrated with the original set of radiomic features to form a comprehensive dataset. \u003c/p\u003e \u003cp\u003e \u003cem\u003eFigure 1: the workflow of Deep-Radiomics Models\u003c/em\u003e \u003c/p\u003e\n\u003ch3\u003eFeature Selection and Validation\u003c/h3\u003e\n\u003cp\u003eTo facilitate a rigorous evaluation of model performance, two distinct datasets were constructed: one comprising only the radiomic features and the other combining both the deep learning-derived features and the radiomic features. These datasets were used for training and evaluating multiple machine-learning models, allowing for a comparative analysis of the impact of incorporating deep learning-derived features on model accuracy.\u003c/p\u003e \u003cp\u003eThe predictive models were trained and validated using imaging data acquired at two critical timepoints: baseline (T1, pre-treatment) and post-first-cycle (T2). This design aligns with the clinical goal of early response prediction, allowing timely therapeutic adjustments before the second chemotherapy cycle.\u003c/p\u003e \u003cp\u003eIn this segment, patients were randomly allocated into training and testing sets, with a 5-fold cross-validation employed to ensure a reliable estimation of model performance. The training set was utilized for the training and refinement of models, while the testing set, isolated from the model fitting process, was used to evaluate the performance of the trained models.\u003c/p\u003e"},{"header":"Results","content":"\u003cdiv id=\"Sec9\" class=\"Section2\"\u003e \u003ch2\u003eAUC values\u003c/h2\u003e \u003cp\u003eWhen using only radiomic features, the AUC values achieved by the RF, XGBoost, LR, and SVM models were 0.85(95%CI: 0.73\u0026ndash;0.97), 0.76 (95%CI: 0.61\u0026ndash;0.91), 0.80 (95%CI: 0.66\u0026ndash;0.94), and 0.59 (95%CI: 0.41\u0026ndash;0.77), respectively shown in Fig.\u0026nbsp;\u003cspan refid=\"Fig1\" class=\"InternalRef\"\u003e2\u003c/span\u003e. Notably, only the SVM model failed to exceed the clinical threshold of 0.70, while the other models demonstrated strong predictive performance, with XGBoost achieving the highest AUC value of 0.84, highlighting its superior capability.\u003c/p\u003e \u003cp\u003e \u003c/p\u003e \u003cp\u003eAfter incorporating deep learning-extracted features with the radiomic features, the AUC values for XGBoost, LR, RF, and SVM increased to 0.92 (95%CI: 0.847-1.000), 0.90 (95%CI: 0.818\u0026ndash;0.982), 0.88 (95%CI: 0.791\u0026ndash;0.969), and 0.61 (95%CI: 0.467\u0026ndash;0.753), respectively, as shown in Fig.\u0026nbsp;\u003cspan refid=\"Fig1\" class=\"InternalRef\"\u003e2\u003c/span\u003e. While the performance of the SVM model remained suboptimal, the AUC values for the other models demonstrated significant improvements, underscoring the value of integrating deep learning-derived features with radiomic features.\u003c/p\u003e \u003cp\u003e \u003cem\u003eStatistical significance of AUC differences between models was assessed using DeLong\u0026rsquo;s test. After Bonferroni correction for multiple comparisons (α\u0026thinsp;=\u0026thinsp;0.05/6 pairwise comparisons), the improvement in XGBoost\u0026rsquo;s AUC from 0.85 (radiomics-only) to 0.92 (radiomics\u0026thinsp;+\u0026thinsp;deep features) was statistically significant (p\u0026thinsp;=\u0026thinsp;0.003). Similarly, LR and RF showed significant improvements (p\u0026thinsp;=\u0026thinsp;0.012 and p\u0026thinsp;=\u0026thinsp;0.021, respectively), while SVM\u0026rsquo;s performance remained unchanged (p\u0026thinsp;=\u0026thinsp;0.42).\u003c/em\u003e \u003c/p\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\u003ePerformance comparison of radiomics-only and combined models using ROC analysis\u003c/p\u003e \u003c/div\u003e \u003c/caption\u003e \u003ccolgroup cols=\"4\"\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c1\" colnum=\"1\"\u003e\u003c/div\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c2\" colnum=\"2\"\u003e\u003c/div\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c3\" colnum=\"3\"\u003e\u003c/div\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c4\" colnum=\"4\"\u003e\u003c/div\u003e \u003cthead\u003e \u003ctr\u003e \u003cth align=\"left\" colname=\"c1\"\u003e\u0026nbsp;\u003c/th\u003e \u003cth align=\"left\" colname=\"c2\"\u003e \u003cp\u003eRadiomics Only (AUC)\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c3\"\u003e \u003cp\u003eCombined (AUC)\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c4\"\u003e \u003cp\u003eDeLong\u0026rsquo;s test (p-value)\u003c/p\u003e \u003c/th\u003e \u003c/tr\u003e \u003c/thead\u003e \u003ctbody\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eXGBoost\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e0.85\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e0.92\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e0.003*\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eLogistic Regression\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e0.80\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e0.90\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e0.012*\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eRandom Forest\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e0.76\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e0.87\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e0.021*\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eSVM\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e0.59\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e0.61\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e0.42\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colspan=\"4\" nameend=\"c4\" namest=\"c1\"\u003e \u003cp\u003e*Note: AUC\u0026thinsp;=\u0026thinsp;Area Under the Curve; * indicates statistical significance (p\u0026thinsp;\u0026lt;\u0026thinsp;0.05).\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\n\u003ch3\u003eDecision Curve Analysis\u003c/h3\u003e\n\u003cp\u003eDecision curve analysis revealed distinct net benefit patterns across threshold probabilities for the three models (Figure X). The XGBoost model achieved the highest net benefit at low thresholds (0.05\u0026ndash;0.30), maintaining stable performance without significant fluctuations. In contrast, logistic regression dominated the mid-threshold range (0.30\u0026ndash;0.55), with a peak net benefit of 0.42 at 0.40, but exhibited declining utility beyond 0.55, reaching negative net benefits at thresholds\u0026thinsp;\u0026gt;\u0026thinsp;0.70. Random forest demonstrated marked variability, with transient peaks at 0.25 (net benefit: 0.38) and 0.50 (0.35), but sharp declines in adjacent ranges (e.g., 0.30\u0026ndash;0.45 and 0.55\u0026ndash;0.65).\u003c/p\u003e \u003cp\u003e \u003c/p\u003e \u003cp\u003eAs baseline references, the \"Treat All\" and \"Treat None\" strategies were outperformed by XGBoost and LR in their respective effective ranges, further demonstrating the potential clinical utility of these models in optimizing treatment decisions.\u003c/p\u003e"},{"header":"Discussion","content":"\u003cp\u003eThis study highlighted the effectiveness of integrating PET/CT imaging data with machine learning models to predict early chemotherapy responses in breast cancer patients. Among the models evaluated, XGBoost demonstrated the highest predictive capability, achieving an Area AUC value of 0.85. This performance exceeded that of RF (AUC: 0.76), LR (AUC: 0.80), and SVM (AUC: 0.59). The superior performance of the XGBoost model underscores its potential as a powerful tool for early chemotherapy response prediction, enabling more precise and personalized treatment strategies for breast cancer patients.\u003c/p\u003e \u003cp\u003eIn this context, XGBoost emerges as a robust solution capable of addressing these limitations. XGBoost is designed to handle large-scale datasets with superior computational efficiency and enhanced resistance to overfitting [\u003cspan citationid=\"CR19\" class=\"CitationRef\"\u003e19\u003c/span\u003e, \u003cspan citationid=\"CR20\" class=\"CitationRef\"\u003e20\u003c/span\u003e]. This algorithm applies gradient boosting principles to iteratively refine prediction accuracy, making it adept at managing complex interactions within high-dimensional data. Therefore, applying XGBoost in predicting breast cancer response should be a key area of focus in current research. Additionally, the application of deep learning models such as the Squeeze-and-Excitation Network (SENet) further enhances the ability to capture significant imaging features from PET/CT data, providing more robust inputs for predictive modeling [\u003cspan citationid=\"CR21\" class=\"CitationRef\"\u003e21\u003c/span\u003e].\u003c/p\u003e \u003cp\u003eBy incorporating deep learning-derived features extracted from PET/CT images using the SENet model, the predictive performance of all machine learning models markedly improved. The AUC values for XGBoost, LR, RF, and SVM increased to 0.92, 0.90, 0.88, and 0.61, respectively. These results underscore the significant value of integrating deep learning-extracted features with traditional radiomic features, enhancing the precision and reliability of treatment planning. While the SVM model continued to underperform compared to other models, the substantial improvement in the XGBoost, LR, and RF models highlights the potential of deep learning-derived features to augment the predictive power of conventional radiomic analysis. The statistically validated superiority of XGBoost underscores its capability to leverage both radiomic and deep learning-derived features, whereas SVM\u0026rsquo;s limitations in handling nonlinear interactions may explain its suboptimal performance.\u003c/p\u003e \u003cp\u003eThe superior performance of the XGBoost model can be largely attributed to its sophisticated algorithmic design, which effectively handles complex interactions among features and exhibits robustness in managing high-dimensional datasets [\u003cspan citationid=\"CR22\" class=\"CitationRef\"\u003e22\u003c/span\u003e]. This model\u0026rsquo;s predictive accuracy is significantly bolstered by its capacity to incorporate and analyze a diverse array of radiomic features. Specifically, the Interquartile Range of the GTV at the pre-treatment stage (T1) and the difference in Mean Standardized Uptake Value (SUVmean) between the pre-treatment stage (T1) and the post-first chemotherapy cycle (T2) emerged as critical predictors. These features likely capture essential variations in tumor metabolism and morphological changes induced by chemotherapy, which are pivotal for assessing treatment efficacy. The detailed and dynamic nature of these predictors contributes to the enhanced reliability and accuracy of the XGBoost model. Furthermore, the model\u0026rsquo;s ability to integrate semi-quantitative PET parameters and clinical features underscores its comprehensive approach to predicting chemotherapy response, thereby reinforcing its potential as a valuable tool in the realm of personalized cancer treatment.\u003c/p\u003e \u003cp\u003eA potential critique of our methodology lies in its deviation from the conventional use of DL models, which are typically employed to predict outcomes directly without explicit feature extraction. In contrast, our approach utilized DL primarily for feature extraction, effectively reframing it into a handcrafted feature model. While this approach may seem unconventional, it enables a more interpretable integration of deep learning-derived features with machine learning models, allowing for a clearer understanding of which features are most predictive of chemotherapy response. Importantly, we speculate that automating the entire DL process\u0026mdash;without a separate feature extraction step\u0026mdash;would not result in a degradation of model performance. In fact, we expect automation to further enhance predictive accuracy and efficiency, particularly as the model continues to evolve with larger datasets and improved optimization techniques. By combining DL-derived features with traditional radiomic analysis, our methodology offers a flexible and scalable framework, capable of adapting to future advancements in medical imaging and DL technologies.\u003c/p\u003e \u003cp\u003eNotably, for DCA curve, although both the XGBoost and Random Forest models outperformed the \"Treat All\" and \"Treat None\" strategies across a substantial portion of the threshold probability range, demonstrating their potential value in clinical decision-making. However, certain limitations of the DCA curves warrant attention. The curves exhibited noticeable fluctuations, likely due to the limited sample size of only 15 patients, which may have introduced instability and amplified the impact of data noise on model performance. Additionally, at lower threshold probabilities (e.g., \u0026lt;\u0026thinsp;0.2) and higher thresholds (\u0026gt;\u0026thinsp;0.5), all curves closely aligned with the \"Treat All\" and \"Treat None\" baselines, indicating that the models offered limited decision-making value under extreme conditions. To address these shortcomings in future studies, we aim to expand the dataset to reduce variability, enhance the stability of the DCA curves, and achieve smoother, more reliable performance metrics.\u003c/p\u003e \u003cp\u003eOur findings are consistent with previous research that underscores the potential of machine learning models in predicting cancer treatment outcomes. Several studies have illustrated how the integration of imaging data with machine learning algorithms can enhance predictive accuracy, thereby improving treatment planning and patient management. For instance, studies by Cai et al. (2023) and Zhang et al. (2018) have demonstrated that radiomic features extracted from imaging data can serve as valuable predictors of treatment response in various cancers [\u003cspan citationid=\"CR7\" class=\"CitationRef\"\u003e7\u003c/span\u003e, \u003cspan citationid=\"CR23\" class=\"CitationRef\"\u003e23\u003c/span\u003e].\u003c/p\u003e \u003cp\u003eThe strength of our approach lies in the comprehensive integration of PET/CT imaging data with machine learning models, which allows for a more nuanced and detailed analysis of tumor characteristics and treatment responses. This methodology leverages the strengths of both imaging and computational analysis, facilitating a deeper understanding of the factors that influence chemotherapy efficacy. However, several potential limitations must be acknowledged. The retrospective design of our study may introduce inherent biases, as the data were collected and analyzed after the fact, which can affect the interpretation of causality and outcome relationships. Additionally, a large sample size would enhance generalizability of our findings, providing a more robust validation of the model's predictive power. The inclusion of a diverse patient population and multi-institutional data could further enhance the reliability and applicability of the results. Future research should focus on prospective studies with larger sample sizes to validate our findings and address these limitations. Moreover, incorporating additional predictive variables, such as genomic data and more comprehensive clinical features, could potentially enhance the models' accuracy and applicability, paving the way for more personalized and effective cancer treatment strategies.\u003c/p\u003e \u003cp\u003eThe ability to predict chemotherapy responses early in the treatment process holds substantial clinical implications. Early identification of patients unlikely to respond to standard chemotherapy regimens enables clinicians to adjust treatment plans promptly, potentially opting for alternative therapies such as radiation or hormone therapy. This personalized approach optimizes treatment efficacy and minimizes unnecessary exposure to ineffective treatments, thereby improving patient outcomes and reducing healthcare costs [\u003cspan citationid=\"CR24\" class=\"CitationRef\"\u003e24\u003c/span\u003e, \u003cspan citationid=\"CR25\" class=\"CitationRef\"\u003e25\u003c/span\u003e]. Early prediction allows for a more tailored therapeutic strategy, ensuring that each patient receives the most appropriate and effective treatment based on their specific tumor characteristics and predicted response.\u003c/p\u003e \u003cp\u003eFuture research should focus on prospective studies to validate the predictive models developed in this study. Additionally, incorporating more predictive variables, such as genomic data and additional radiomic features, could enhance the models' accuracy and applicability. Testing these models in real-world clinical settings will also be crucial to assess their practical utility and refine them based on clinical feedback.\u003c/p\u003e"},{"header":"Conclusion","content":"\u003cp\u003eIn conclusion, our study demonstrated that integrating PET/CT imaging data with machine learning models, including XGBoost, SVM, LR, and RF model, effectively predicts early chemotherapy responses in breast cancer patients. The XGBoost model showed superior predictive capability, facilitating personalized treatment strategies. Early identification of non-responders allows for timely interventional therapy, enhancing treatment efficacy and avoiding unnecessary continuation of ineffective treatments. This approach underscores the potential of advanced machine learning models in improving patient outcomes and optimizing cancer treatment management.\u003c/p\u003e"},{"header":"Declarations","content":"\u003cp\u003e\u003cstrong\u003eFunding\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eOO\u0026ndash; Supported by Purdue University Research Funding.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eHuman Ethics and Consent to Participate declarations\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eThe imaging data employed in this study were sourced from the QIN-Breast dataset available through The Cancer Imaging Archive (TCIA). This dataset is publicly accessible and has been de-identified in compliance with the Health Insurance Portability and Accountability Act (HIPAA) regulations, ensuring the protection of patient privacy. As the data are anonymized and publicly available, the study was exempt from institutional review board (IRB) approval. Nonetheless, we adhered to all relevant guidelines and regulations governing the use of such data. In accordance with TCIA\u0026apos;s data usage policies, we have cited the dataset appropriately and have not attempted to re-identify any individual participants.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eCompeting interests\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eThe authors declare that they have no competing interests. \u0026nbsp;\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eAuthor Contribution\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eZ.J. conceptualized the study, designed the methodology, performed radiomic analysis, developed machine learning models, and wrote the original manuscript. J.L. contributed to data collection, preprocessing, and preliminary statistical validation. C.H., Y.Y., and C.N. provided technical resources and reviewed the manuscript. O.O. supervised the project, provided critical revisions to the manuscript, and approved the final version. All authors read and approved the final manuscript.\u003c/p\u003e"},{"header":"References","content":"\u003col\u003e\u003cli\u003e\u003cspan\u003eSung H et al. Global cancer statistics 2020: GLOBOCAN estimates of incidence and mortality worldwide for 36 cancers in 185 countries. 2021. 71(3): pp. 209\u0026ndash;49.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eŁukasiewicz S et al. Breast cancer\u0026mdash;epidemiology, risk factors, classification, prognostic markers, and current treatment strategies\u0026mdash;an updated review. 2021. 13(17): p. 4287.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eFang S et al. A real-world clinicopathological model for predicting pathological complete response to neoadjuvant chemotherapy in breast cancer. 2024. 14: p. 1323226.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eHadebe B et al. The role of PET/CT in breast cancer. 2023. 13(4): p. 597.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eLi T et al. Applications of FAPI PET/CT in the diagnosis and treatment of breast and the most common gynecologic malignancies: a literature review. 2024. 14: p. 1358070.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eIqbal MJ, et al. Clin Appl Artif Intell Mach Learn cancer diagnosis: Look into future. 2021;21(1):270.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eZhang B, Shi H. and H.J.J.o.m.h. Wang, \u003cem\u003eMachine learning and AI in cancer prognosis, prediction, and treatment selection: a critical approach.\u003c/em\u003e 2023: pp. 1779\u0026ndash;1791.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eRasool A et al. Improved machine learning-based predictive models for breast cancer diagnosis. 2022. 19(6): p. 3211.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eZhang Y et al. PET radiomics in lung cancer: advances and translational challenges. 2024. 11(1): p. 81.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eZhou Y et al. Integrating 18F-FDG PET/CT radiomics and body composition for enhanced prognostic assessment in patients with esophageal cancer. 2024. 24(1): p. 1402.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eMenon N et al. Performance of radiomics-based artificial intelligence systems in the diagnosis and prediction of treatment response and survival in esophageal cancer: a systematic review and meta-analysis of diagnostic accuracy. 2023. 36(6): p. doad034.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eSim Y et al. Multiparametric MRI\u0026ndash;based radiomics model for predicting human papillomavirus status in oropharyngeal squamous cell carcinoma: optimization using oversampling and machine learning techniques. 2024. 34(5): pp. 3102\u0026ndash;12.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eLohmann P et al. \u003cem\u003eRadiomics in radiation oncology\u0026mdash;basics, methods, and limitations.\u003c/em\u003e 2020. 196(10): pp. 848\u0026ndash;855.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eMali SA et al. Making radiomics more reproducible across scanner and imaging protocol variations: a review of harmonization methods. 2021. 11(9): p. 842.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eLiu Y et al. Multimodality deep learning radiomics predicts pathological response after neoadjuvant chemoradiotherapy for esophageal squamous cell carcinoma. 2024. 15(1): p. 277.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eWei W et al. Deep learning radiomics for prediction of axillary lymph node metastasis in patients with clinical stage T1\u0026ndash;2 breast cancer. 2023. 13(8): p. 4995.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eLi X et al. Data QIN-breast 2016. 5.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eWahl RL et al. \u003cem\u003eFrom RECIST to PERCIST: evolving considerations for PET response criteria in solid tumors.\u003c/em\u003e 2009. 50(Suppl 1): p. 122S-150S.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eZhang X et al. \u003cem\u003ePredicting missing values in medical data via XGBoost regression.\u003c/em\u003e 2020. 4: pp. 383\u0026ndash;394.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eXu Y et al. Predicting patients' satisfaction with doctors in online medical communities: An approach based on XGBoost algorithm. 2022. 34(4): pp. 1\u0026ndash;17.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eHu J, Shen L, Sun G. \u003cem\u003eSqueeze-and-excitation networks\u003c/em\u003e. in \u003cem\u003eProceedings of the IEEE conference on computer vision and pattern recognition\u003c/em\u003e. 2018.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eMa B, et al. XGBLC: improved survival prediction model based XGBoost. 2022;38(2):410\u0026ndash;8.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eCai J et al. A radiomics model for predicting the response to bevacizumab in brain necrosis after radiotherapy. 2020. 26(20): pp. 5438\u0026ndash;47.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eRomeo V et al. Assessment and prediction of response to neoadjuvant chemotherapy in breast cancer: A comparison of imaging modalities and future perspectives. 2021. 13(14): p. 3521.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eLiang X, et al. Early prediction of pathological complete response to neoadjuvant chemotherapy combining DCE-MRI and apparent diffusion coefficient values. Breast Cancer. 2022;22(1):1250.\u003c/span\u003e\u003c/li\u003e\u003c/ol\u003e"}],"fulltextSource":"","fullText":"","funders":[],"hasAdminPriorityOnWorkflow":false,"hasManuscriptDocX":true,"hasOptedInToPreprint":true,"hasPassedJournalQc":"","hasAnyPriority":false,"hideJournal":false,"highlight":"","institution":"","isAcceptedByJournal":true,"isAuthorSuppliedPdf":false,"isDeskRejected":"","isHiddenFromSearch":false,"isInQc":false,"isInWorkflow":false,"isPdf":false,"isPdfUpToDate":true,"isWithdrawnOrRetracted":false,"journal":{"display":true,"email":"[email protected]","identity":"medical-oncology","isNatureJournal":false,"hasQc":true,"allowDirectSubmit":false,"externalIdentity":"medo","sideBox":"Learn more about [Medical Oncology](https://www.springer.com/journal/12032)","snPcode":"12032","submissionUrl":"https://submission.nature.com/new-submission/12032/3","title":"Medical Oncology","twitterHandle":"","acdcEnabled":true,"dfaEnabled":true,"editorialSystem":"stoa","reportingPortfolio":"Springer Hybrid","inReviewEnabled":true,"inReviewRevisionsEnabled":false},"keywords":"","lastPublishedDoi":"10.21203/rs.3.rs-6474899/v1","lastPublishedDoiUrl":"https://doi.org/10.21203/rs.3.rs-6474899/v1","license":{"name":"CC BY 4.0","url":"https://creativecommons.org/licenses/by/4.0/"},"manuscriptAbstract":"\u003ch2\u003eBackground\u003c/h2\u003e \u003cp\u003eEnhancing the accuracy of tumor response predictions enables the development of tailored therapeutic strategies for patients with breast cancer. In this study, we developed deep-radiomic models to enhance the prediction of chemotherapy response after the first treatment cycle.\u003c/p\u003e\u003ch2\u003eMethods\u003c/h2\u003e \u003cp\u003e18F- Fludeoxyglucose (FDG) PET/CT imaging data and clinical record from 60 breast cancer patients were retrospectively obtained from the Cancer Imaging Archive (QIN-BREAST). PET/CT scans were conducted at three distinct stages of treatment; prior to the initiation of chemotherapy (T1), following the first cycle of chemotherapy (T2), and after the full chemotherapy regimen (T3). The patient's primary gross tumor volume (GTV) was delineated on PET images using a 40% threshold of the maximum standardized uptake value (SUVmax). Radiomic features were extracted from the GTV based on the PET/CT images. In addition, a Squeeze-and-Excitation Network (SENet) deep learning model was employed to generate additional features from the PET/CT images for combined analysis. A XGBoost machine learning model was developed and compared with the conventional machine learning (ML) algorithm (random forest [RF], logistic regression [LR] and support vector machine [SVM]). The performance of each model was assessed using receiver operating characteristics area under the curve (ROC AUC) analysis, and prediction accuracy in a validation cohort. Model performance was evaluated through 5-fold cross-validation on the entire cohort, with data splits stratified by treatment response categories to ensure balanced representation.\u003c/p\u003e\u003ch2\u003eResults\u003c/h2\u003e \u003cp\u003eThe AUC values for the machine learning models using only radiomic features were 0.85(XGBoost), 0.76 (RF), 0.80 (LR), and 0.59 (SVM), with XGBoost showing the best performance. After incorporating additional deep learning-derived features from SENet, the AUC values increased to 0.92, 0.88, 0.90, and 0.61, respectively, demonstrating significant improvements in predictive accuracy. Predictions were based on pre-treatment (T1) and post-first-cycle (T2) imaging data, enabling early assessment of chemotherapy response after the initial treatment cycle.\u003c/p\u003e\u003ch2\u003eConclusion\u003c/h2\u003e \u003cp\u003eIntegrating deep learning-derived features significantly enhanced the performance of predictive models for chemotherapy response in breast cancer patients. This study demonstrated the superior predictive capability of the XGBoost model, emphasizing its potential to optimize personalized therapeutic strategies by accurately identifying patients unlikely to respond to chemotherapy after the first treatment cycle.\u003c/p\u003e","manuscriptTitle":"18F-FDG PET/CT-Based Deep Radiomic Models for Enhancing Chemotherapy Response Prediction in Breast Cancer","msid":"","msnumber":"","nonDraftVersions":[{"code":1,"date":"2025-05-15 10:35:17","doi":"10.21203/rs.3.rs-6474899/v1","editorialEvents":[{"type":"communityComments","content":0},{"type":"decision","content":"Revision requested","date":"2025-07-13T00:07:53+00:00","index":"","fulltext":""},{"type":"editorInvitedReview","content":"","date":"2025-06-30T20:39:28+00:00","index":"hide","fulltext":""},{"type":"reviewerAgreed","content":"285961458848171424110478067969697715442","date":"2025-06-21T05:33:14+00:00","index":"hide","fulltext":""},{"type":"reviewerAgreed","content":"248661589068439950228394870670798352327","date":"2025-06-20T18:44:22+00:00","index":"hide","fulltext":""},{"type":"editorInvitedReview","content":"","date":"2025-05-20T08:39:31+00:00","index":"hide","fulltext":""},{"type":"reviewerAgreed","content":"211602614727140800998419555757702803830","date":"2025-05-12T13:24:27+00:00","index":"hide","fulltext":""},{"type":"reviewersInvited","content":"","date":"2025-05-12T02:17:19+00:00","index":"","fulltext":""},{"type":"editorAssigned","content":"","date":"2025-04-18T17:59:16+00:00","index":"","fulltext":""},{"type":"checksComplete","content":"","date":"2025-04-18T17:58:49+00:00","index":"","fulltext":""},{"type":"submitted","content":"Medical Oncology","date":"2025-04-18T00:11:24+00:00","index":"","fulltext":""}],"status":"published","journal":{"display":true,"email":"[email protected]","identity":"medical-oncology","isNatureJournal":false,"hasQc":true,"allowDirectSubmit":false,"externalIdentity":"medo","sideBox":"Learn more about [Medical Oncology](https://www.springer.com/journal/12032)","snPcode":"12032","submissionUrl":"https://submission.nature.com/new-submission/12032/3","title":"Medical Oncology","twitterHandle":"","acdcEnabled":true,"dfaEnabled":true,"editorialSystem":"stoa","reportingPortfolio":"Springer Hybrid","inReviewEnabled":true,"inReviewRevisionsEnabled":false}}],"origin":"","ownerIdentity":"65635f7f-74b6-4723-947f-db3853d486d9","owner":[],"postedDate":"May 15th, 2025","published":true,"recentEditorialEvents":[],"rejectedJournal":[],"revision":"","amendment":"","status":"published-in-journal","subjectAreas":[],"tags":[],"updatedAt":"2025-08-18T16:00:52+00:00","versionOfRecord":{"articleIdentity":"rs-6474899","link":"https://doi.org/10.1007/s12032-025-02982-0","journal":{"identity":"medical-oncology","isVorOnly":false,"title":"Medical Oncology"},"publishedOn":"2025-08-11 15:57:29","publishedOnDateReadable":"August 11th, 2025"},"versionCreatedAt":"2025-05-15 10:35:17","video":"","vorDoi":"10.1007/s12032-025-02982-0","vorDoiUrl":"https://doi.org/10.1007/s12032-025-02982-0","workflowStages":[]},"version":"v1","identity":"rs-6474899","journalConfig":"researchsquare"},"__N_SSP":true},"page":"/article/[identity]/[[...version]]","query":{"redirect":"/article/rs-6474899","identity":"rs-6474899","version":["v1"]},"buildId":"8U1c8b4HqxoKbykW_rLl7","isFallback":false,"isExperimentalCompile":false,"dynamicIds":[84888],"gssp":true,"scriptLoader":[]}