Development and external validation of a machine learning-based multimodal radiomics nomogram for predicting progression-free survival in triple-negative breast cancer

Introduction

Breast cancer ranked as the second most commonly diagnosed malignancy and the fourth leading cause of cancer-related mortality worldwide. Its considerable prevalence, coupled with a growing trend toward earlier onset and elevated mortality, poses a substantial threat to women’s health globally (1,2). Among subtypes, the TNBC—characterized by the absence of estrogen receptor (ER), progesterone receptor (PR), and human epidermal growth factor receptor 2 (HER2) expression—is particularly aggressive, with higher recurrence rates, reduced survival, and a lack of targeted therapy options, leading to markedly worse outcomes (3). Medical imaging has expanded beyond diagnosis to support treatment guidance and prognostic prediction. This progression is enabled by modern digital techniques that capture detailed, multidimensional data on tumor heterogeneity. The concept of comprehensive imaging analysis is often traced back to Lambin et al. (4); in 2012, they established that radiomics enables high-throughput extraction and quantitative analysis of medical imaging features, providing more comprehensive diagnostic information than traditional techniques (5). This approach is primarily applied in oncology for tumor characterization, differential diagnosis, treatment response evaluation, and prognostic prediction (6,7). The field has been further accelerated by integrating machine learning and deep learning into computer-aided diagnosis (CAD) systems, facilitating the development of robust predictive models for clinical outcomes. Numerous studies have validated the efficacy of these artificial intelligence (AI)-driven approaches in evaluating treatment response and predicting prognosis, particularly for neoadjuvant chemotherapy outcomes and long-term survival in breast cancer patients (8-10).

Despite these advances, most existing predictive models for TNBC rely on single-modality data [e.g., magnetic resonance imaging (MRI) alone] and focus exclusively on the tumor core while ignoring the peritumoral microenvironment. For instance, Kim et al. developed an MRI-based radiomics model for disease-free survival (DFS) but did not incorporate digital breast tomosynthesis (DBT) or peritumoral features (11). Moreover, most existing machine learning models lack external validation and multimodal integration. Consequently, a key knowledge gap remains: the absence of a standardized, externally validated framework that combines DBT, multiparametric MRI, and peritumoral radiomics for progression-free survival (PFS) prediction in TNBC. To address this gap, we aimed to develop and validate models for predicting PFS by integrating clinicopathological parameters, conventional DBT and MRI features, and radiomics from both the tumor and peritumoral regions. This study developed a clinicopathological-conventional imaging model and multiple radiomics models based on diverse MRI sequences and DBT projections. The integration of these models facilitated the identification of independent risk factors for PFS. A resulting nomogram provides a quantitative, non-invasive preoperative tool to predict recurrence and outcomes in TNBC, supporting personalized treatment planning. We present this article in accordance with the TRIPOD reporting checklist (available at https://tcr.amegroups.com/article/view/10.21037/tcr-2026-1-0341/rc).

Methods

Study sample

No formal sample size calculation was performed due to the exploratory nature of the analysis; the sample was based on availability of eligible patients during the study period. This study was conducted in accordance with the Declaration of Helsinki and its subsequent amendments. The study was approved by the Ethics Committees of The First Affiliated Hospital of Kunming Medical University (2024L No.17) and The Third Affiliated Hospital of Kunming Medical University (No. KYLX2024-185). Informed consent was waived in this retrospective study. From an initial 140 pathologically confirmed TNBC cases at two tertiary centers (Center 1, The First Affiliated Hospital of Kunming Medical University; Center 2, The Third Affiliated Hospital of Kunming Medical University), 108 patients met the inclusion criteria and were included in the final analysis (75 from Center 1 and 33 from Center 2). The patient selection process is detailed in Figure 1. With 15 progression events in the training set and 7 in the validation set, the study is limited by low event numbers, increasing the risk of model overfitting and unstable performance estimates. Inclusion criteria: (I) pathologically confirmed first-time unilateral TNBC with complete immunohistochemistry results and no history of other tumors; (II) completion of DBT and MRI examinations within 2 weeks prior to the invasive procedure; and (III) availability of complete clinical data and regular follow-up. Exclusion criteria: (I) history of antitumor therapy prior to imaging examination; (II) incomplete imaging sequences; (III) poor image quality affecting evaluation and delineation; and (IV) distant metastasis before surgery.

Figure 1 The inclusion and exclusion process. Center 1, The First Affiliated Hospital of Kunming Medical University; Center 2, The Third Affiliated Hospital of Kunming Medical University. DBT, digital breast tomosynthesis; MRI, magnetic resonance imaging; TNBC, triple-negative breast cancer.

Breast images from DBT and MRI were assessed by two board-certified breast radiologists (>10 years of experience) using the Breast Imaging Reporting and Data System (BI-RADS) atlas [2013] and quantitative parameters [e.g., apparent diffusion coefficient (ADC), maximal diameter] were averaged from their independent measurements, with discrepancies resolved by a third senior specialist (imaging signs of the breast, Figures 2-9). All patients received treatment according to institutional guidelines, including surgery, adjuvant chemotherapy, and radiotherapy. PFS was defined as the time from the date of pathological diagnosis to the date of first documented disease progression (local recurrence, distant metastasis, or contralateral breast cancer) or death from any cause, whichever occurred first. Outcome assessment was performed by clinicians who were blinded to the radiomics and predictor data.

Figure 2 DBT manifestations: (A) mass; (B) mass with fine calcification; (C) structural distortion. DBT, digital breast tomosynthesis.

Figure 3 Lesion shapes: (A) oval; (B) lobulated; (C) irregular.

Figure 4 Lesion margins: (A) smooth; (B) not smooth; (C) spiculation sign.

Figure 5 Lesion T2 signals: (A) iso-level signal; (B) slightly elevated signal; (C) elevated signal; (D) mixed signal.

Figure 6 Types of edema around breast lesions: (A) perifocal edema; (B) skin edema; (C) edema in front of the pectoral muscles.

Figure 7 Internal enhancement characteristics of the lesion: (A) uniform; (B) non-uniform; (C) ring-shaped enhancement.

Figure 8 The conditions of adjacent structures to the lesion: (A) nipple retraction/invasion; (B) skin retraction/involvement; (C) pectoral muscle involvement.

Figure 9 Different characteristics of two sequences: (A) peripheral vascular sign; (B) DWI annular high signal. DWI, diffusion-weighted imaging.

Imaging

DBT

In Center 1, mammographic examinations were performed using the MAMMOMAT Inspiration system (Siemens Healthineers, Erlangen, Germany) with automatic exposure control (AEC) mode. Standard craniocaudal (CC) and mediolateral oblique (MLO) projections were acquired for each breast. In Center 2, mammography was conducted using the Selenia Dimensions system (Hologic, Inc., Marlborough, MA, USA), also utilizing AEC mode, with bilateral CC and MLO views obtained according to standard mammographic protocols.

MRI

Both centers performed breast MRI with dedicated coils in the prone position. Center 1 used a 3.0-T Philips scanner with the following parameters: T1-weighted imaging (T1WI) (TR/TE 400/10 ms); fat-suppressed T2-weighted imaging (T2WI) (5,000/60 ms); diffusion-weighted imaging (DWI) (3,300/71 ms, b=0,1000 s/mm²); and dynamic contrast-enhanced (DCE)-MRI (six 60-second dynamic phases). Center 2 used a 1.5-T Siemens MAGNETOM scanner with: T1WI (4.43/1.39 ms); T2WI (5,600/56 ms); DWI (4,900/84 ms, b=0,800 s/mm²); and DCE-MRI (six 60-second dynamic phases). Common parameters included: slice thickness 4 mm (T2WI, DWI) to 1 mm (T1WI, DCE-MRI), with interslice gaps of 0.2–2.0 mm; field of view 350 mm × 350 mm (Center 1) or 360 mm × 360 mm (Center 2). Both protocols administered contrast agent (gadopentetate dimeglumine at Center 1; gadodiamide at Center 2) at 0.2 mmol/kg via power injector (2.0 mL/s) followed by 20 mL saline flush.

Image region of interest (ROI) delineation

ROI delineation was performed by radiologists blinded to the clinical outcomes and pathologic data. DBT (CC/MLO) and multiparametric MRI sequences were imported into 3D Slicer 5.2.2 for analysis. For each case, the largest confirmed lesion was manually delineated layer-by-layer to include all affected areas. The tumor ROI was expanded outward by 5 and 10 mm using the ‘Hollow’ tool, and non-parenchymal areas were excluded with the ‘Erase’ function. Five ROIs were defined: ROI 0 (tumor); ROI 1/3 (tumor + 10 mm/5 mm); ROI 2/4 (10 mm/5 mm peritumoral).

Image feature extraction and dimensionality reduction

To mitigate inter-scanner variability, images were preprocessed by standardization (μ±3σ), grayscale discretization, and resampling to 1.0 mm³ isotropic voxels. A total of 851 radiomic features (162 first-order, 14 shape, 675 texture) were then extracted via the PyRadiomics module in 3D Slicer. Following feature import into Python 3.7 and labeling (1 for progression, 0 for non-progression), features were standardized. Non-significant features (two-sample t-test, P≥0.05) were excluded, and dimensionality was further reduced using least absolute shrinkage and selection operator (LASSO) regression (appendices available at https://cdn.amegroups.cn/static/public/tcr-2026-1-0341-1.pdf).

Statistical analysis

Statistical analyses were conducted using SPSS 26.0 and R (version 4.2.1; http://www.r-project.org). Normally distributed continuous variables are expressed as mean ± standard deviation, with comparisons between groups performed using the two-sample t-test. Categorical data were analyzed using the Chi-squared test. Univariate and multivariate Cox regression analyses were employed to identify independent risk factors. Kaplan-Meier survival curves, along with log-rank tests, were used for survival analysis. Receiver operating characteristic (ROC) curves were constructed, and the predictive performance of each model was assessed based on the area under the curve (AUC), sensitivity (Se), specificity (Sp), and accuracy (Acc). Calibration curves, clinical decision curves, and the concordance index (C-index) were used to evaluate the clinical applicability of the nomogram. Statistical significance was set at P<0.05. Patients were stratified into high- and low-risk groups based on the optimal cut-off value (e.g., median) of the Rad score for Kaplan-Meier analysis. There were no missing data for the predictors included in the final analysis.

Prediction model construction

The model was developed using 75 patients from Center 1 as the training set (with 50% cross-validation) and externally validated on 33 independent patients from Center 2. Five machine learning algorithms [AdaBoost (AB), light gradient boosting machine (LGBM), logistic regression (LR), random forest (RF), and extreme gradient boosting (XGB)] were applied to build predictive models. Rad scores were subsequently computed from the multimodal radiomic features using logistic regression. For external validation, the Rad scores for patients in Center 2 were calculated using the same formula derived from the LASSO regression in the training set (appendices available at https://cdn.amegroups.cn/static/public/tcr-2026-1-0341-2.pdf).

The following Rad scores were calculated: Rad score 0, Rad score 1, Rad score 2, Rad score 3, and Rad score 4. Clinicopathological and conventional imaging features, along with Rad scores derived from multimodal omics data, were analyzed using univariate and multivariate Cox regression. Variables with a P value <0.10 in the univariate analysis were included in the multivariate analysis and further assessed via backward stepwise regression to identify independent risk factors for PFS and construct a nomogram.

Results

Patient characteristics

The study included 108 patients with triple-negative breast cancer (TNBC), with a mean age of 50.42 years (range, 25–79 years) and a median follow-up of 56 months (range, 12–86 months) as of August 31, 2024. The cohort was divided into a training cohort (n=75; 15 with disease progression) and an external validation cohort (n=33; 7 with disease progression). There were no significant differences in baseline characteristics between the development and validation cohorts (Table 1), confirming the comparability of the two datasets.

Prediction model construction

In the training set (n=75), 15 patients (20%) experienced disease progression, while in the external validation set (n=33), 7 patients (21%) had events. The baseline characteristics of the 108 included patients were analyzed using both univariate and multivariate Cox proportional hazards regression models. The results of the univariate analysis are presented in Table 2. Survival analysis was conducted using Kaplan-Meier curves and log-rank tests. Five machine learning algorithms (AB, LGBM, LR, RF, and XGB) were employed to develop predictive models for TNBC, based on tumor core, peritumoral region, and combined features, utilizing both single-sequence and multimodal imaging data. Single-sequence radiomics models derived from DBT and MRI were constructed using various machine learning classifiers to predict PFS in TNBC patients. The predictive performance of the models varied across different imaging sequences. Among the MRI-based models, the XGB model demonstrated the best performance, while the RF model outperformed the others in DBT (Table 3).

Rad scores for each patient were calculated according to the established Rad score formula. Patients were then stratified into high-risk and low-risk groups, with the median value serving as the cutoff point. Survival analysis was performed using Kaplan-Meier curves and log-rank tests. Rad score_0 corresponds to the tumor body, Rad score_1 to the tumor body plus a 10 mm peritumoral region, Rad score_2 to the 10 mm peritumoral region, Rad score_3 to the tumor body plus a 5mm peritumoral region, and Rad score_4 to the 5 mm peritumoral region. Kaplan-Meier survival curves revealed a significant difference between the two groups (P<0.01). (Figures 10,11).

Figure 10 DBT-Rad score KM-curve. 0, no PFS event; 1, PFS event occurred. DBT, digital breast tomosynthesis; KM, Kaplan-Meier; PFS, progression-free survival.

Figure 11 MRI-Rad score KM-curve. 0, no PFS event; 1, PFS event occurred. KM, Kaplan-Meier; MRI, magnetic resonance imaging; PFS, progression-free survival.

Nomogram construction

Univariate and multivariate Cox regression analyses were performed on clinicopathological-radiological features and Rad scores derived from multimodal omics data. Variables with a P value <0.10 in the univariate analysis were subsequently included in the multivariate analysis, which was conducted using backward stepwise regression. This analysis identified perilesional edema, World Health Organization (WHO) grade, lymphovascular invasion, DBT-Rad score_2 (10 mm peritumoral region), and MRI Rad score_0 (tumor) as independent prognostic factors for PFS in TNBC patients (P<0.05). The final multivariable Cox regression model for the nomogram included the following predictors with their respective coefficients and hazard ratios (HRs): perilesional edema (HR =2.08), WHO grade (HR =4.08), lymphovascular invasion (HR =4.49), DBT-Rad score_2 (10 mm peritumoral region) (HR =3.86), MRI Rad score_0 (tumor) (HR =4.31). A nomogram was then developed to incorporate these significant predictors. The nomogram provides a visual representation of the scoring system for these predictive factors, allowing the calculation of a total score for individual patients and thereby estimating their probability of 2-, 3-, and 5-year PFS. The nomogram (Figure 12) demonstrated good predictive performance for PFS in TNBC patients, with AUC values of 0.858 [95% confidence interval (CI): 0.788–0.928] in the training cohort and 0.736 (95% CI: 0.676–0.796) in the external validation cohort. The calibration curve and decision curve analysis further confirmed the nomogram’s significant clinical utility (Figure 13). To use the nomogram, locate the patient’s value on each predictor axis, draw a vertical line to the ‘Points’ axis to obtain a score for each predictor, sum these scores to obtain a ‘Total points’, and then draw a vertical line down to the ‘2-/3-/5-year PFS’ axes to estimate the survival probability.

Figure 12 The nomogram for predicting PFS in patients with TNBC. Perilesional edema: 0, absent; 1, present. Lymphovascular invasion: 0, absent; 1, present. DBT, digital breast tomosynthesis; MRI, magnetic resonance imaging; PFS, progression-free survival; TNBC, triple-negative breast cancer; WHO, World Health Organization.

Figure 13 Performance of the nomogram for predicting PFS in TNBC patients in the (A) training cohort and (B) external validation cohort. The first line is the AUC of the nomogram. The second line is the calibration curve of the nomogram for predicting PFS in TNBC patients. The third line is the decision curve of the nomogram for predicting PFS in TNBC patients. AUC, area under the curve; PFS, progression-free survival; TNBC, triple-negative breast cancer.

Discussion

Building upon established prognostic factors in TNBC (12-14), this study aimed to investigate the association of radiomic features extracted from MRI and DBT, along with clinicopathological and conventional imaging factors, with PFS in patients with TNBC. By integrating clinical data cohorts with radiomic features extracted from different imaging modalities, this research systematically evaluates multiple machine learning algorithms to develop predictive models for disease progression. This study proposes an innovative framework that integrates multimodal image analysis and AI to enhance prognostic stratification in the management of TNBC. This approach may establish a new paradigm for personalized outcome prediction in breast cancer. Radiomic features were extracted from multiparametric MRI sequences (including T1WI, T2WI, DCE-MRI, DWI, and ADC) and dual-view DBT (CC and MLO).

This study presents the first comprehensive radiomics assessment integrating multiparametric MRI with DBT for breast cancer evaluation. Notably, the 5mm peritumoral ADC-based model demonstrated exceptional predictive performance for PFS (training AUC =0.91; validation AUC =0.81). The superior performance of the ADC-based model likely stems from its ability to quantify water diffusion restriction, where lower ADC values indicate higher cellularity and tumor aggressiveness. This finding aligns with prior evidence demonstrating the efficacy of ADC radiomics in predicting pathological complete response to neoadjuvant chemotherapy, with reported AUCs of 0.87 (training) and 0.85 (validation) (15). The limited sample size presents a potential overfitting risk. Nevertheless, this study underscores the value of ADC maps, which offer critical insights into tissue cellularity and the tumor microenvironment (TME). These findings support ADC maps as a significant biomarker of tumor aggressiveness.

Regarding peritumoral radiomics, the 10-mm margin consistently outperformed the 5-mm margin. This may be explained by the fact that the TME—including cancer-associated fibroblasts, immune infiltrates, and inflamed stroma—often extends beyond 5 mm in aggressive TNBC. Indeed, Jiang et al. demonstrated that peritumoral heterogeneity captured by radiomics is significantly associated with immune suppression and unfavorable prognosis in TNBC, supporting the biological rationale for evaluating a wider peritumoral region (16). Furthermore, previous studies on breast cancer molecular subtyping have identified optimal peritumoral margins of 6 mm for specific classification tasks, exceeding the conventional 5-mm boundary (17). Therefore, the 10-mm peritumoral margin may more accurately reflect the TME and its underlying biological behavior. However, the current lack of a standardized peritumoral definition, with selections varying across studies, limits result comparability. Future studies with larger cohorts are warranted to define the optimal peritumoral margin, which could enhance the Acc of radiomics-based predictive models.

Compared with conventional clinical and histopathological parameters, the radiomics score offers distinct advantages in non-invasive detection, early diagnosis, and therapeutic efficacy assessment. It provides more accurate, individualized predictions, thereby supporting more informed clinical decision-making (18-20). In this study, multimodal radiomics features were quantified into a Rad score using a specific calculation formula. The Rad score is a numerical value that facilitates the visualization of complex imaging characteristics and aids in predicting PFS in patients with TNBC. This score was derived from multiple breast MRI sequences and DBT images, including CC and MLO views, with varying ROIs. By integrating multiple imaging features, the Rad score provides a comprehensive quantitative profile of tumor characteristics, serving as an effective clinical tool for assessing prognosis and recurrence risk.

Radiomics feature analysis demonstrates considerable promise in multiple aspects of breast cancer management, including lesion characterization, staging, subtyping, and prognostic prediction (21,22). This study developed a nomogram integrating clinicopathological features, conventional imaging, and the Rad score. As a practical prognostic tool widely used in oncology, this nomogram synthesizes independent predictors into a visual format, thereby enabling comprehensive survival analysis and supporting clinical decision-making (23-25). In this study, the Rad score was calculated using multimodal radiomics features derived from both the tumor core and peritumoral region across multiple MRI and DBT sequences. The findings revealed that perilesional edema, WHO grade, lymphovascular invasion, DBT peritumoral 10mm Rad score, and MRI tumor Rad score were independent predictors of PFS in TNBC patients. Based on these factors, we developed a nomogram to visualize the prediction model. The nomogram exhibited excellent predictive performance, with AUC values of 0.858 (training cohort) and 0.736 (external validation cohort), indicating its clinical utility. Previous studies have shown that a combined model integrating MRI (T1WI, T2WI) radiomics features with clinicopathological parameters achieved superior predictive performance for DFS in TNBC patients (AUC =0.844) compared to a clinicopathological-only model (AUC =0.765) (11). Notably, this study advances the field by demonstrating higher predictive Acc.

This study has several limitations. First, the small sample size (n=108) and low event numbers (15 training, 7 validation) are major concerns. With only seven events in the validation set, the 95% CI for the validation AUC (0.52–0.95) is extremely wide, indicating poor precision. Current consensus recommends at least 100 events for stable external validation (TRIPOD). Therefore, our results should be considered hypothesis-generating and preliminary. Second, this study has limitations including its retrospective design, which may introduce selection bias, and potential feature instability due to the sample size. In future research, expanding the sample size could enable the exploration of additional machine learning and deep learning methods, potentially enhancing the predictive performance. Additionally, optimizing the parameters of the selected model through techniques such as grid search, random search, or Bayesian optimization may help identify the optimal parameter combination, thereby improving the model’s stability and generalizability.

Conclusions

In conclusion, the nomogram-integrating perilesional edema, WHO grade, lymphovascular invasion, DBT peritumoral 10-mm Rad score, and MRI tumor Rad score, effectively predicts recurrence risk and quantifies 2-, 3-, and 5-year PFS probabilities in TNBC patients, providing a valuable tool for guiding clinical decisions and optimizing follow-up strategies. But this study developed and externally validated a machine learning-based multimodal radiomics nomogram that integrates clinicopathological factors, conventional imaging features, and peritumoral radiomics from DBT and MRI to predict PFS in TNBC. The nomogram shows promising discrimination and clinical utility. However, due to the limited sample size, low event numbers, and wide CIs, these findings should be considered preliminary. Prospective validation in larger, multi-center cohorts is required before clinical implementation.

Acknowledgments

None.

Footnote

Reporting Checklist: The authors have completed the TRIPOD reporting checklist. Available at https://tcr.amegroups.com/article/view/10.21037/tcr-2026-1-0341/rc

Data Sharing Statement: Available at https://tcr.amegroups.com/article/view/10.21037/tcr-2026-1-0341/dss

Peer Review File: Available at https://tcr.amegroups.com/article/view/10.21037/tcr-2026-1-0341/prf

Funding: This work was supported by the Joint Foundation of Yunnan Provincial Science and Technology Department (No. 202101AY070001-101), the First-Class Discipline Team of Kunming Medical University (No. 2024XKTDTS03), and the Kunming Medical University-level Innovation and Entrepreneurship Training Program for College Students (No. PX-425207).

Conflicts of Interest: All authors have completed the ICMJE uniform disclosure form (available at https://tcr.amegroups.com/article/view/10.21037/tcr-2026-1-0341/coif). The authors have no conflicts of interest to declare.

Ethical Statement: The authors are accountable for all aspects of the work in ensuring that questions related to the accuracy or integrity of any part of the work are appropriately investigated and resolved. This study was conducted in accordance with the Declaration of Helsinki and its subsequent amendments. The study was approved by the Ethics Committees of The First Affiliated Hospital of Kunming Medical University (2024L No.17) and The Third Affiliated Hospital of Kunming Medical University (No. KYLX2024-185). Informed consent was waived in this retrospective study.

Open Access Statement: This is an Open Access article distributed in accordance with the Creative Commons Attribution-NonCommercial-NoDerivs 4.0 International License (CC BY-NC-ND 4.0), which permits the non-commercial replication and distribution of the article with the strict proviso that no changes or edits are made and the original work is properly cited (including links to both the formal publication through the relevant DOI and the license). See: https://creativecommons.org/licenses/by-nc-nd/4.0/.

References

1. Global cancer burden growing, amidst mounting need for services. Saudi Med J 2024;45:326-7.
2. Dvir K, Giordano S, Leone JP. Immunotherapy in Breast Cancer. Int J Mol Sci 2024;25:7517. [Crossref] [PubMed]
3. Adrada BE, Moseley TW, Kapoor MM, et al. Triple-Negative Breast Cancer: Histopathologic Features, Genomics, and Treatment. Radiographics 2023;43:e230034. [Crossref] [PubMed]
4. Lambin P, Rios-Velazquez E, Leijenaar R, et al. Radiomics: extracting more information from medical images using advanced feature analysis. Eur J Cancer 2012;48:441-6. [Crossref] [PubMed]
5. Guiot J, Vaidyanathan A, Deprez L, et al. A review in radiomics: Making personalized medicine a reality via routine imaging. Med Res Rev 2022;42:426-40. [Crossref] [PubMed]
6. Tokuda Y, Suzuki Y, Oda S, et al. A Radiogenomic Model using MRI and Gene Signature to Predict Complete Response in Breast Cancer. Eur J Radiol 2026;200:112824. [Crossref] [PubMed]
7. Shao J, Wei M, Li K, et al. Radiomics: Current Applications and Future Directions. MedComm (2020) 2026;7:e70773.
8. Sampa MB, Abdul Aziz NH, Rahman MS, et al. Reinforcement learning for medical image analysis: a systematic review of algorithms, engineering challenges, and clinical deployment. Comput Assist Surg (Abingdon) 2026;31:2597553. [Crossref] [PubMed]
9. Sabry M, Balaha HM, Ali KM, et al. AI-Driven Breast Cancer Diagnosis: A Systematic Review of Imaging Modalities, Deep Learning, and Explainability. Cancers (Basel) 2026;18:1305. [Crossref] [PubMed]
10. Wang Z, Ma T, Li Y, et al. Multidimensional MRI radiomics-based model predicts recurrence risk in triple-negative breast cancer. NPJ Precis Oncol 2026; Epub ahead of print. [Crossref]
11. Kim S, Kim MJ, Kim EK, et al. MRI Radiomic Features: Association with Disease-Free Survival in Patients with Triple-Negative Breast Cancer. Sci Rep 2020;10:3750. [Crossref] [PubMed]
12. Yuan C, Xu G, Zhan X, et al. Molybdenum target mammography-based prediction model for metastasis of axillary sentinel lymph node in early-stage breast cancer. Medicine (Baltimore) 2023;102:e35672. [Crossref] [PubMed]
13. Pintican R, Fechete R, Boca B, et al. Predicting the Early Response to Neoadjuvant Therapy with Breast MR Morphological, Functional and Relaxometry Features-A Pilot Study. Cancers (Basel) 2022;14:5866. [Crossref] [PubMed]
14. Yang S, Wang Z, Wang C, et al. Comparative Evaluation of Machine Learning Models for Subtyping Triple-Negative Breast Cancer: A Deep Learning-Based Multi-Omics Data Integration Approach. J Cancer 2024;15:3943-57. [Crossref] [PubMed]
15. Zhou Z, Adrada BE, Candelaria RP, et al. Predicting pathological complete response to neoadjuvant systemic therapy for triple-negative breast cancers using deep learning on multiparametric MRIs. Annu Int Conf IEEE Eng Med Biol Soc 2023;2023:1-4. [Crossref] [PubMed]
16. Jiang L, You C, Xiao Y, et al. Radiogenomic analysis reveals tumor heterogeneity of triple-negative breast cancer. Cell Rep Med 2022;3:100694. [Crossref] [PubMed]
17. Han Y, Huang M, Xie L, et al. The value of intratumoral and peritumoral radiomics features based on multiparametric MRI for predicting molecular staging of breast cancer. Front Oncol 2025;15:1379048. [Crossref] [PubMed]
18. Wu X, Dai W. Beyond Imaging: Integrating Radiomics, Genomics, and Multi-Omics for Precision Breast Cancer Management. Cancers (Basel) 2025;17:3408. [Crossref] [PubMed]
19. Cheng C, Wang Y, Zhao J, et al. Deep Learning and Radiomics in Triple-Negative Breast Cancer: Predicting Long-Term Prognosis and Clinical Outcomes. J Multidiscip Healthc 2025;18:319-27. [Crossref] [PubMed]
20. Zhang J, Wu Q, Lei P, et al. MRI-based radiomics models for early predicting pathological response to neoadjuvant chemotherapy in triple-negative breast cancer: A systematic review and meta-analysis. J Appl Clin Med Phys 2025;26:e70296. [Crossref] [PubMed]
21. Zhang J, Wu Q, Lei P, et al. MRI-based radiomics models for early predicting pathological response to neoadjuvant chemotherapy in triple-negative breast cancer: A systematic review and meta-analysis. J Appl Clin Med Phys 2025;26:e70296. [Crossref] [PubMed]
22. Ferro A, Bottosso M, Dieci MV, et al. Clinical applications of radiomics and deep learning in breast and lung cancer: A narrative literature review on current evidence and future perspectives. Crit Rev Oncol Hematol 2024;203:104479. [Crossref] [PubMed]
23. Wang Y, Zhang J, Wu Y, et al. Postoperative radiotherapy for ductal carcinoma in situ: survival prediction and clinical decision support using a nomogram-based approach. Sci Rep 2025;16:433. [Crossref] [PubMed]
24. Liu YX, Liu QH, Hu QH, et al. Ultrasound-Based Deep Learning Radiomics Nomogram for Tumor and Axillary Lymph Node Status Prediction After Neoadjuvant Chemotherapy. Acad Radiol 2025;32:12-23. [Crossref] [PubMed]
25. Wenwen Jiang Z. Integrating ultrasound radiomics and clinicopathological features for machine learning-based survival prediction in patients with nonmetastatic triple-negative breast cancer. BMC Cancer 2025;25:291. [Crossref] [PubMed]