Imagine a world where doctors could predict, before starting treatment, whether breast cancer patients will fully respond to chemotherapy—potentially saving lives and sparing unnecessary suffering. This groundbreaking study harnesses the power of advanced imaging and AI to do just that, offering hope in the fight against one of the world's deadliest diseases. But here's where it gets controversial: Is relying on machine learning to make such critical medical decisions the future of personalized care, or does it risk oversimplifying the complex biology of cancer? Stick around, and this is the part most people miss—the subtle interplay between early and peak phases of MRI scans that could revolutionize how we tackle breast cancer treatment.
Predicting Breast Cancer Outcomes to Neoadjuvant Therapy Through the Fusion of Radiomic and Deep-Learning Insights from Early and Peak Stages of Dynamic Contrast-Enhanced Magnetic Resonance Imaging
Investigation
Freely Accessible (https://www.springernature.com/gp/open-science/about/the-fundamentals-of-open-access-and-open-research)
Released: November 11, 2025
*Yabin Zhang¹ ,² co-first author,
*Jiumei Cai³ co-first author,
*Chunxiao Cui⁴,
*Shouliang Qi¹ ,² &
*Dan Zhao³
BMC Cancer (https://bmccancer.biomedcentral.com/) edition 25, Article 1747 (2025) Reference this article
Summary
Non-invasive forecasting of pathological complete response (pCR) in breast cancer patients receiving neoadjuvant therapy (NAT) holds immense value for refining surgical plans and tailoring treatment approaches. In this research, we crafted a predictive framework that merges conventional radiomics with 3D deep learning attributes drawn from the early and peak phases of dynamic contrast-enhanced magnetic resonance imaging (DCE-MRI) to anticipate pCR following NAT.
Approach
Our retrospective analysis encompassed 234 breast cancer patients from two medical centers. Dataset 1 (comprising 204 individuals) served for model creation, while Dataset 2 (with 30 participants) underwent external validation. We extracted conventional radiomics and 3D deep learning characteristics from the entire DCE-MRI scan and the tumor's region of interest (ROI). These features were combined, then refined through independent sample t-tests and least absolute shrinkage and selection operator (LASSO) regression, ultimately selecting the top ten most influential features for model development. Logistic regression facilitated the construction of predictive models, with performance assessed via receiver operating characteristic curves and area under the curve (AUC). The DeLong method evaluated distinctions in AUC values across models, and SHAP (SHapley Additive exPlanations) explored the interplay between model features and pCR.
Findings
Among models relying solely on conventional radiomics traits, the integrated approach combining early and peak phases of DCE-MRI yielded the strongest pCR forecasting. This model's efficacy was further elevated by incorporating 3D deep learning elements. The premier model, RDEP, fusing radiomics and deep learning data from both early and peak phases, attained AUC scores of 0.892 (95% CI: 0.853–0.922) on Dataset 1 and 0.825 (95% CI: 0.713–0.886) on Dataset 2. DeLong testing indicated notable statistical disparities between RDEP and alternative predictive models (p < 0.05). SHAP scrutiny revealed that two radiomics texture attributes were especially pivotal in driving the model's decisions.
Conclusion
Blending traditional radiomics with 3D deep learning features across multiple phases of DCE-MRI enables precise prediction of pCR in response to NAT for breast cancer. Utilizing multi-phase imaging alongside varied features boosts forecasting precision, and the developed model could guide individualized therapeutic strategies.
Expert Assessments (https://bmccancer.biomedcentral.com/articles/10.1186/s12885-025-15095-8/peer-review)
Introduction
Breast cancer stands as a leading global malignancy and a primary contributor to cancer-related fatalities among women [1, 2]. Neoadjuvant therapy (NAT) represents a vital modality for managing early-stage, high-risk, and locally advanced breast cancer. It facilitates tumor shrinkage, reduces Tumor-Node-Metastasis (TNM) staging, enhances surgical success rates, boosts prospects for breast-preserving operations, curbs distant metastasis, and informs future therapeutic choices [3,4,5,6]. Certain patients attain pathological complete response (pCR) via NAT, and in contrast to those without pCR, they enjoy markedly superior long-term survival rates [7]. Consequently, anticipating pCR in breast cancer patients becomes essential for crafting customized treatment regimens and enhancing patient prognosis [8]. For beginners, think of pCR as the 'gold standard' outcome in cancer treatment—it's when no cancer cells remain detectable under a microscope after therapy, signaling a potentially curative response.
In clinical settings, imaging techniques form the cornerstone for evaluating NAT effectiveness in breast cancer patients preoperatively, encompassing mammography, ultrasound, magnetic resonance imaging (MRI), diffuse optical spectroscopy, and Positron Emission Tomography–Computed Tomography (PET-CT) [9,10,11,12]. Among these, breast MRI emerges as the premier non-invasive tool for pCR prediction, closely mirroring pathological assessments [10,11,13,14]. Specifically, dynamic contrast-enhanced MRI (DCE-MRI) vividly delineates tumor contours, morphology, structure, and internal blood flow dynamics [15,16]. These elements are vital for the quantitative evaluation of tumors. With radiomics advancements, extensive quantitative analysis of medical imaging data extracts myriad quantitative features, applicable to disease identification, prognosis forecasting, and treatment efficacy assessment [17,18,19,20], offering substantial clinical utility. To illustrate for newcomers, radiomics is like turning a medical image into a treasure map of data points—each pixel or voxel reveals hidden patterns about tumor behavior, much like how a detective analyzes clues at a crime scene.
Radiomics excels in forecasting NAT outcomes in breast cancer patients [21,22,23,24,25]. For instance, Eun et al. [22] leveraged specialized software for texture scrutiny on patient MRI scans, developing a pCR predictive model rooted in texture features. Zheng et al. [24] incorporated features from diverse breast tissue zones to anticipate pCR, underscoring the potential influence of the tumor microenvironment on NAT success. Sutton et al. [26] deployed automated radiomics techniques, including semi-automatic tumor delineation, to construct a recursive feature elimination-random forest classifier for pCR prediction. Yet, in these endeavors, most features stemmed from bespoke extraction tools like Pyradiomics [27], TexRAD [28], IBEX [29], MaZda [30], and LIFEx [31], with these features predominantly being manually designed. Conversely, deep learning algorithms autonomously derive intricate high-level features from medical images, uncovering tumor details that conventional radiomics might miss, and they've been employed in breast cancer detection, classification, and outcome prediction [32,33]. Li et al. [34] showcased the superiority of a hybrid model merging traditional radiomics and deep learning features in forecasting axillary lymph node involvement in breast cancer. Peng et al. [35] utilized deep learning approaches to predict pathological responses to NAT in breast cancer, discovering these outperformed traditional radiomics methods. However, their deep learning features were sourced solely from 2D tumor slices, inevitably overlooking valuable 3D spatial details. Three-dimensional feature extraction grasps more comprehensive spatial tumor information, enabling richer analysis. Integrating 3D deep learning features with conventional radiomics for pCR prediction in breast cancer patients is an innovative and demanding avenue of exploration.
Radiomics models grounded in DCE-MRI assess NAT effectiveness in breast cancer patients [21,22,23,24,25,36], forecast HER2 and Ki-67 expression levels [37,38], predict sentinel lymph node involvement [39,40], and classify molecular tumor subtypes [41,42]. Nevertheless, these investigations extracted tumor features exclusively from the early [22] or peak [23] phases of dynamic contrast-enhanced imaging. Huang et al. [42] integrated different DCE-MRI phases to categorize breast cancer molecular subtypes, yielding encouraging outcomes. This underscores the potential clinical value and predictive power of multi-phase DCE-MRI. To date, no definitive consensus exists on which DCE-MRI phase delivers optimal pCR predictive performance.
This investigation proposes harnessing baseline DCE-MRI early and peak phase scans, extracting and amalgamating radiomics and 3D deep learning features, and formulating machine learning models to forecast pCR in response to NAT in breast cancer. By evaluating the efficacy of diverse predictive models, the aim is to pinpoint the most effective feature combinations for NAT outcome prediction in patients. This could empower clinicians to devise more precise, individualized treatment plans for breast cancer sufferers. The study's primary novelties and contributions encompass: (a) A consolidated model that precisely forecasts neoadjuvant therapy response in breast cancer using DCE-MRI. (b) Methodologically, radiomics and 3D deep learning features from early and peak DCE-MRI phases are examined, clarified, and fused. (c) A suite of imaging biomarkers is established, predictive of neoadjuvant chemotherapy response in breast cancer. (d) External dataset validation confirms the robust generalizability of the constructed predictive model.
Materials and Methods
Patient Selection and Data
This retrospective investigation drew from data across two institutions (Institution A: Liaoning Cancer Hospital, Institution B: Affiliated Hospital of Qingdao University). The project received endorsement from both centers' medical ethics boards (Ref. 20221101; QYFY WZLL 28631), with waived patient consent. We retrospectively compiled clinical records of female breast cancer patients diagnosed pathologically and undergoing neoadjuvant therapy prior to surgery, spanning stages I to III, devoid of distant metastasis and prior radiation, surgery, or other breast cancer interventions.
Inclusion criteria included: (a) Diagnosis of primary unilateral breast cancer, with either a solitary lesion or accompanied by satellite foci (MRI-confirmed before NAT), lacking distant spread. (b) Completion of at least four NAT cycles. (c) Subsequent surgical intervention post-NAT. (d) Availability of comprehensive pathological and pre-NAT biopsy details. (e) Invasive ductal carcinoma histology. Exclusion criteria encompassed: (a) Bilateral breast cancer. (b) History of distant metastasis or alternative malignancies. (c) Incomplete pathological, clinical, or imaging records.
Adhering to criteria, Institution A enrolled 204 cases (Dataset 1), and Institution B contributed 30 (Dataset 2). Dataset 1 cases were partitioned into training and testing subsets at a 7:3 ratio, with Dataset 2 serving as external validation. Patient enrollment flow is illustrated in Figure 1.
Flowchart of case inclusion, exclusion, and data grouping
Full size image (https://bmccancer.biomedcentral.com/articles/10.1186/s12885-025-15095-8/figures/1)
Figure 2 depicts the overall experimental workflow. Initially, DCE-MRI data were procured and preprocessed. Subsequently, conventional radiomics and 3D deep learning features were derived from early and peak phase DCE-MRI scans. Next, feature refinement occurred via independent sample t-tests and LASSO regression. Lastly, models were built and validated.
Experimental workflow, encompassing data acquisition and preprocessing, feature extraction, refinement, model development, and assessment
Full size image (https://bmccancer.biomedcentral.com/articles/10.1186/s12885-025-15095-8/figures/2)
MRI Acquisition and Tumor Delineation
All participants received contrast-enhanced breast MRI prior to neoadjuvant therapy. For DCE-MRI, Institution A employed a 1.5T GE Signa MRI system, while Institution B utilized 3.0T SIEMENS Skyra and 3.0T GE Discovery MRI units. Following mask acquisition, contrast was administered intravenously at 0.1 mmol/kg via high-pressure injector. Institution A used Omniscan (gadodiamide, GE Healthcare, Ireland), and Institution B opted for Magnevist (gadopentetate dimeglumine, Bayer AG, Germany). Post-injection, the 1.5T GE Signa at Institution A captured 8 successive non-overlapping scans, each 58 seconds, with 3.5mm slices, 0mm spacing. The 3.0T SIEMENS Skyra at Institution B acquired 5 successive non-overlapping scans, each 90 seconds, with 1mm slices, 0.2mm spacing. The 3.0T GE Discovery at Institution B recorded 5 successive non-overlapping scans, each 65 seconds, with 1mm slices, 0.5mm spacing.
Pre-NAT DCE-MRI scans were retrieved from Picture Archiving and Communication Systems, exported in DICOM format. Guided by each case's DCE-MRI time-signal intensity curve, early and peak phase sequence images were exported. Two radiologists (one with 3 years, the other over 5 years of breast imaging expertise) delineated tumor ROIs on each axial transverse slice using ITK-SNAP (www.itk-snap.org, Version 4.0.1), then auto-generated the full lesion ROI. A senior radiologist (over 15 years experience) reviewed the ROIs. Inter-reader agreement was gauged via Dice similarity coefficient, averaging 0.836 (95% CI: 0.802–0.867), denoting solid delineation reliability.
Image Preprocessing
Preprocessing entailed N4 bias field adjustment [43] and ROI-based cropping. Outcomes are displayed in Figure 3.
Image preprocessing involving N4 bias field correction and image cropping. (a) Preprocessing steps; (b) N4 bias field correction applied to lesion regions
Full size image (https://bmccancer.biomedcentral.com/articles/10.1186/s12885-025-15095-8/figures/3)
Bias field artifacts—low-frequency intensity distortions common in MRI—can skew image quality, impacting analysis. N4 bias field correction, an algorithm for medical imaging, rectifies these to enhance accuracy. For instance, imagine a photo that's too bright in one corner; this tool evens it out like adjusting lighting in a room.
Original scans measured 512×512×48 voxels, with lesions occupying minimal space. To streamline feature extraction, ROI-based cropping was applied. Using MONAI toolkit (https://monai.io/, Version 1.2.0), images were cropped per ROI, masks resized accordingly. To mitigate inter-center magnetic field and protocol variances, uniform resampling and z-score standardization were applied post-N4 correction, resampling to a consistent grid via trilinear interpolation for volumes and nearest-neighbor for masks. Z-score normalization used per-volume effective voxels (non-zero), scaling intensities. For multi-phase DCE-MRI, normalization was phase-specific to preserve dynamic signals.
Feature Extraction
Features were extracted from the full 3D image (whole-image) and ROI (tumor-focused).
Conventional Radiomics Extraction: Employing Pyradiomics (https://pyradiomics.readthedocs.io/en/latest/index.html, version 3.0.1), we derived 1,409 conventional radiomics features per phase, including First Order, Shape, GLCM, GLSZM, GLRLM, NGTDM, and GLDM. For beginners, these are like statistical summaries of the image: First Order looks at basic pixel intensities, Shape describes the tumor's form, while matrices like GLCM analyze how pixel values relate spatially, revealing texture patterns that might indicate tumor aggressiveness.
3D Deep Learning Extraction: A pre-trained 3D-ResNet-18 network (https://github.com/pytorch/vision/blob/main/torchvision/models/video/resnet.py) extracted 512 3D deep learning features per phase. This convolutional neural network, adapted for 3D data, captures complex spatial and semantic tumor traits. Its first layer was adjusted for single-channel input (vs. RGB), weights averaged across channels, and the classification head removed for feature extraction. Scans were resized and normalized; ROI masks ensured lesion-focused features.
Feature Refinement
With features numbering initially in the thousands, dimensionality reduction was key to avert overfitting—the 'curse of dimensionality,' where too many variables confuse the model, like trying to track too many suspects in a mystery novel. Training-set-only refinement began with independent t-tests identifying features with p < 0.05 differences between pCR and non-pCR groups. LASSO with 10-fold cross-validation shrank coefficients, discarding irrelevant features. Top 10 features per phase/type were selected for modeling.
Six feature sets were defined:
R_E: Radiomics from Early phase;
R_P: Radiomics from Peak phase;
R_EP: Radiomics from Early and Peak phases;
RD_E: Radiomics and Deep Learning from Early phase;
RD_P: Radiomics and Deep Learning from Peak phase;
RD_EP: Radiomics and Deep Learning from Early and Peak phases.
Model Development and Assessment
Logistic regression (Scikit-Learn, Version 1.2.2) built six models: RE, RP, REP, RDE, RDP, RDEP. Hyperparameters were tuned via grid search with 10-fold cross-validation. SMOTE balanced imbalanced data by generating synthetic minority samples, improving minority class recognition—like creating extra practice examples for a rare scenario in training data.
Dataset 1 trained and tested models; Dataset 2 validated generalizability. Metrics included specificity, sensitivity, accuracy, AUC. ROC curves and matrices visualized performance. DeLong tested AUC differences; SHAP explained feature contributions [47]. 95% CIs used 2,000 bootstrap samples.
Results
Patient Demographics
Table 1 outlines patient and clinical traits for Datasets 1 and 2. In Dataset 1 (n=204), pCR (n=53) and non-pCR (n=151) groups differed only in Ki-67 status (p=0.028). Dataset 2 (n=30) showed no group differences.
Full size table (https://bmccancer.biomedcentral.com/articles/10.1186/s12885-025-15095-8/tables/1)
Feature Insights
Post-integration of early/peak radiomics and deep learning (3,842 features), t-tests reduced to 489 significant ones (236 early: 212 radiomics, 14 deep; 263 peak: 233 radiomics, 30 deep). LASSO narrowed to 65 (29 early: 23 radiomics, 6 deep; 36 peak: 28 radiomics, 8 deep). Top 10 features were chosen: 7 early (6 radiomics, 1 deep), 3 peak (2 radiomics, 1 deep). Figure 5 shows LASSO MSE and coefficient shifts with Lambda. Figure 6 compares these features' means, deviations, and distributions between groups. Higher pCR means prevailed except for two features.
LASSO MSE and feature coefficients evolving with Lambda
Full size image (https://bmccancer.biomedcentral.com/articles/10.1186/s12885-025-15095-8/figures/5)
Top 10 features comparison: (a) Means and deviations; (b) Data spread
Full size image (https://bmccancer.biomedcentral.com/articles/10.1186/s12885-025-15095-8/figures/6)
Radiomics-Only Models' Efficacy
Figure 7(a,b) presents ROC curves for radiomics-only models on whole-image vs. ROI-based features. ROI-based models outperformed, with AUC ranges 0.689–0.803 vs. 0.685–0.743. Table 2 details metrics; R_EP excelled on ROI (AUC 0.803, 95% CI: 0.712–0.851).
ROC curves for all models: (a) Whole-image radiomics-only; (b) ROI-based radiomics-only; (c) Whole-image radiomics+deep; (d) ROI-based radiomics+deep
Full size table (https://bmccancer.biomedcentral.com/articles/10.1186/s12885-025-15095-8/tables/2)
Radiomics+Deep Learning Models' Efficacy
Incorporating deep learning boosted performance (Figure 7(c,d); Table 3). RDEP led on ROI (AUC 0.892, 95% CI: 0.853–0.922). DeLong p-values in Figure 8 showed RDEP significantly superior (all p<0.05).
Full size table (https://bmccancer.biomedcentral.com/articles/10.1186/s12885-025-15095-8/tables/3)
DeLong p-values across ROI-based models
Full size image (https://bmccancer.biomedcentral.com/articles/10.1186/s12885-025-15095-8/figures/8)
External Validation Efficacy
Focusing on superior ROI-based models, Table 4 and Figure 9 confirm strong generalizability; RD_EP achieved 0.825 AUC (95% CI: 0.713–0.886).
Full size table (https://bmccancer.biomedcentral.com/articles/10.1186/s12885-025-15095-8/tables/4)
ROC curves on external validation: (a) Radiomics-only; (b) Radiomics+deep
Full size image (https://bmccancer.biomedcentral.com/articles/10.1186/s12885-025-15095-8/figures/9)
Discussion
PCR Prediction's Clinical Relevance
Breast cancer ranks among women's top health threats [1,2,3]. NAT shrinks tumors, easing surgery and improving survival [48]. pCR signals excellent prognosis, but non-pCR patients endure futile treatments and side effects [49]. Predicting pCR via imaging could personalize care, avoiding burdens. For instance, imagine sparing a patient chemotherapy that won't work, freeing them for alternative therapies.
Feature Refinement and Key Attributes
Initial features overwhelmed; t-tests and LASSO streamlined them. SHAP (Figure 10) highlighted GLSZM and GLRLM features' importance—GLSZM quantifies gray level zones for texture uniformity, GLRLM captures 3D structural heterogeneity. Higher values in pCR groups suggest complexity linked to markers like ER/PR/Ki-67 [48,50,51]. But here's the controversial angle: Could these features oversimplify tumor biology, ignoring environmental factors or patient genetics? Deep learning features, though opaque, added value [35,54]. DCA (Figure 11) reinforced RD_EP's clinical edge.
Importance of different features via SHAP analysis
Full size image (https://bmccancer.biomedcentral.com/articles/10.1186/s12885-025-15095-8/figures/10)
DCA: (a) Radiomics-only; (b) Radiomics+deep
Full size image (https://bmccancer.biomedcentral.com/articles/10.1186/s12885-025-15095-8/figures/11)
Early vs. Peak Phases
Early-phase models often outperformed peak-phase, capturing perfusion dynamics [59,60]. Yet, peak-phase holds visualization strengths. This sparks debate: Is early-phase universally superior, or context-dependent?
Methodological Strengths
Table 5 compares our multi-phase, radiomics+deep approach to prior works. We expanded beyond single-phase [23] or 2D deep learning [61], yielding superior accuracy.
Full size table (https://bmccancer.biomedcentral.com/articles/10.1186/s12885-025-15095-8/tables/5)
Limitations and Prospects
Small datasets limit scope; larger, diverse cohorts needed. Reliance on MRI overlooks multimodal data like PET-CT [62]. Feature fusion was basic; advanced methods could enhance. Current 78.7% accuracy misses some patients—critical for high-stakes decisions. Future: Integrate clinical data, refine networks, boost accuracy.
Conclusions
Merging radiomics with 3D deep learning across DCE-MRI phases accurately predicts pCR, guiding personalized breast cancer care. This paves research and application paths.
Data Accessibility
Data available from authors upon request.
Abbreviations
AUC: Area Under the Curve; CI: Confidence Interval; DCE-MRI: Dynamic Contrast-Enhanced MRI; ER: Estrogen Receptor; GLCM: Gray Level Co-occurrence Matrix; GLDM: Gray Level Dependence Matrix; GLSZM: Gray Level Size Zone Matrix; GLRLM: Gray Level Run Length Matrix; HER-2: Human Epidermal Growth Factor Receptor 2; LASSO: Least Absolute Shrinkage and Selection Operator; LR: Logistic Regression; MRI: Magnetic Resonance Imaging; MSE: Mean Squared Error; NAT: Neoadjuvant Therapy; NGTDM: Neighboring Gray Tone Difference Matrix; non-pCR: Non-Pathological Complete Response; pCR: Pathological Complete Response; PET-CT: Positron Emission Tomography–Computed Tomography; PR: Progesterone Receptor; ResNet: Residual Neural Network; ROC: Receiver Operating Characteristic; ROI: Region of Interest; SHAP: SHapley Additive exPlanations; SMOTE: Synthetic Minority Over-sampling Technique; TNM: Tumor-Node-Metastasis
References
[1] Sung H, et al. Global Cancer Statistics 2020: GLOBOCAN Estimates... CA Cancer J Clin. 2021;71(3):209–49.
[2] Wilkinson L, Gathani T. Understanding Breast Cancer as a Global Health Concern. Br J Radiol. 2022;95(1130):20211033.
[3] Gradishar WJ, et al. Breast Cancer, Version 4.2017... J Natl Compr Canc Netw. 2018;16(3):310–20.
[4] The Society of Breast Cancer China Anti-Cancer Association... China Oncol. 2023;33(12):1092–187.
[5] Derks MG, van de Velde CJ. Neoadjuvant Chemotherapy in Breast Cancer... Lancet Oncol. 2018;19(1):2–3.
[6] Thompson AM, Moulder-Thompson SL. Neoadjuvant Treatment of Breast Cancer. Ann Oncol. 2012;23:x231–6.
[7] Kong X, et al. Meta-Analysis Confirms Achieving Pathological Complete Response... Eur J Cancer. 2011;47(14):2084–90.
[8] Montemurro F, et al. Neoadjuvant or Adjuvant Chemotherapy in Early Breast Cancer? Expert Opin Pharmacother. 2020;21(9):1071–82.
[9] Tran WT, et al. Predicting Breast Cancer Response... Br J Cancer. 2017;116(10):1329–39.
[10] Dialani V, et al. Role of Imaging in Neoadjuvant Therapy for Breast Cancer. Ann Surg Oncol. 2015;22:1416–24.
[11] Li H, et al. MRI and PET/CT for Evaluation... Breast. 2018;40:106–15.
[12] Urso L, et al. PET-Derived Radiomics and Artificial Intelligence... International Journal of Molecular Sciences. 2022;23(21):13409.
[13] Berg WA, et al. Diagnostic Accuracy of Mammography... Radiology. 2004;233(3):830–49.
[14] Yeh E, et al. Prospective Comparison of Mammography... Am J Roentgenol. 2005;184(3):868–77.
[15] Pinker K, et al. The Potential of Multiparametric MRI of the Breast. Br J Radiol. 2017;90(1069):20160715.
[16] Mann RM, et al. Breast MRI: State of the Art. Radiology. 2019;292(3):520–36.
[17] Lambin P, et al. Radiomics: Extracting More Information... Eur J Cancer. 2012;48(4):441–6.
[18] Gillies RJ, et al. Radiomics: Images Are More Than Pictures... Radiology. 2016;278(2):563–77.
[19] Aerts HJ. The Potential of Radiomic-Based Phenotyping... JAMA Oncol. 2016;2(12):1636–42.
[20] Liu Z, et al. The Applications of Radiomics... Theranostics. 2019;9(5):1303.
[21] Zheng G, et al. Prediction of Neoadjuvant Chemotherapy Pathological Complete Response... BMC Med Imaging. 2024;24(1):22.
[22] Eun NL, et al. Texture Analysis with 3.0-T MRI... Radiology. 2020;294(1):31–41.
[23] Liu Z, et al. Radiomics of Multiparametric MRI... Clin Cancer Res. 2019;25(12):3538–47.
[24] Zheng G, et al. Predicting Pathological Complete Response... J Cancer Res Clin Oncol. 2024;150(3):147.
[25] Sutton EJ, et al. A Machine Learning Model... Breast Cancer Res. 2020;22:1–11.
[26] van Griethuysen JJ, et al. Computational Radiomics System... Cancer Res. 2017;77(21):e104–7.
[27] Szczypiński PM, et al. MaZda—a Software Package... Comput Methods Programs Biomed. 2009;94(1):66–76.
[28] LHS MD, et al. TexRad-Feedback plc-Cancer Management Imaging Software. 2015.
[29] Zhang L, et al. IBEX: An Open Infrastructure Software... Med Phys. 2015;42(3):1341–53.
[30] Nioche C, et al. LIFEx: A Freeware for Radiomic Feature... Cancer Res. 2018;78(16):4786–9.
[31] Abhisheka B, et al. A Comprehensive Review... Arch Comput Methods Eng. 2023;30(8):5023–52.
[32] Li X, et al. Comparison of Traditional Radiomics... Acad Radiol. 2023;30(7):1281–7.
[33] Peng Y, et al. Pretreatment DCE-MRI-Based Deep Learning... Front Oncol. 2022;12:846775.
[34] Li W, et al. Predicting Breast Cancer Response... NPJ Breast Cancer. 2020;6(1):63.
[35] Li C, et al. Intratumoral and Peritumoral Radiomics... J Magn Reson Imaging. 2021;54(3):703–14.
[36] Feng S, et al. Radiomics of Dynamic Contrast-Enhanced... Front Oncol. 2022;12:847880.
[37] Ma M, et al. A Radiomics Model for Preoperative Predicting Sentinel Lymph Node Metastasis... Front Oncol. 2022;12:884599.
[38] Liu J, et al. Radiomics Analysis of Dynamic Contrast-Enhanced... Front Oncol. 2019;9:980.
[39] Huang G, et al. Molecular Subtypes of Breast Cancer Identified... Insights into Imaging. 2024;15(1):127.
[40] Tustison NJ, et al. N4ITK: Improved N3 Bias Correction. IEEE Trans Med Imaging. 2010;29(6):1310–20.
[41] Köppen M. The Curse of Dimensionality. In 5th Online World Conference... 2000;1:4–8.
[42] Chawla NV, et al. SMOTE: Synthetic Minority Over-sampling Technique. J Artif Intell Res. 2002;16:321–57.
[43] Lundberg SM, Lee SI. A Unified Approach to Interpreting Model Predictions. Adv Neural Inf Process Syst. 2017;30.
[44] Wang H. Evaluation of the Efficacy of Neoadjuvant Chemotherapy for Breast Cancer. Drug Design, Development and Therapy. 2020;14:2423–2433.
[45] Oba T, et al. Neoadjuvant Chemotherapy-Induced Decrease... BMC Cancer. 2020;20:1–10.
[46] Turashvili G, Brogi E. Tumor Heterogeneity in Breast Cancer. Front Med. 2017;4:227.
[47] Huang Y, et al. Longitudinal MRI-Based Fusion Novel Model... EClinicalMedicine. 2023;58:101899.
[48] Hussain J. Deep Learning Black Box Problem. 2019.
[49] Yu Y, et al. Radiomic Model Based on Magnetic Resonance Imaging... Translational Oncology. 2024;46:101985.
[50] Jiang C, et al. The Prediction of pCR and Chemosensitivity... Translational Oncology. 2024;46:101985.
[51] Pesapane F, et al. Radiomics of MRI for the Prediction... Cancers. 2021;13(17):4271.
[52] Braman NM, et al. Intratumoral and Peritumoral Radiomics... Breast Cancer Res. 2017;19:1–14.
[53] Cain EH, et al. Multivariate Machine Learning Models... Breast Cancer Res Treat. 2019;173:455–63.
[54] Fan M, et al. Radiomics of Tumor Heterogeneity... Frontiers in Molecular Biosciences. 2021;8:622219.
[55] Romeo V, et al. Tumor Segmentation Analysis... European Journal of Radiology. 2020;126:108907.
[56] Macura KJ, et al. Patterns of Enhancement... Radiographics. 2006;26(6):1719–34.
[57] Ebrahimi A, et al. Introducing Transfer Learning... In 2020 35th International Conference... IEEE.
[58] Li Y, et al. Deep Learning Radiomic Analysis... Frontiers in Oncology. 2023;12:1041142.
[59] Urso L, et al. ¹⁸F-FDG PET/CT Radiomic Analysis... La Radiologia Medica. 2025;130(4):543–554.
So, what's your take? Do you think AI-driven predictions like this will fully transform breast cancer care, or might they introduce new ethical dilemmas around trust and over-reliance? Share your thoughts in the comments—do you agree this multi-phase approach is a game-changer, or disagree that we should prioritize simpler methods? Let's discuss!
