Artificial Intelligence for Breast Cancer Molecular Subtype Prediction From Medical Imaging: A Systematic Review of Literature

Abstract

Background

Breast cancer molecular subtypes (Luminal A, luminal B, HER2-positive, and triple-negative) are typically determined through biopsy and immunohistochemistry, procedures that are invasive and prone to interpretative variability. Artificial intelligence applied to medical imaging has emerged as a non-invasive alternative to support subtype prediction. However, methodological rigor, risk of bias and clinical translatability of existing studies remain insufficiently characterized.

Methods

We conducted a systematic review in accordance with the PRISMA 2020 guidelines. Searches were performed in Scopus, Web of Science, ScienceDirect, and PubMed to identify English-language, peer-reviewed original articles published between January 1, 2020, and March 25, 2025. Two reviewers independently screened titles and abstracts using Rayyan, retrieved potentially eligible full-text articles, and assessed risk of bias and applicability using an AI-adapted QUADAS framework based on QUADAS-2.

Results

From 626 identified records, 76 full-text articles were assessed for eligibility. The AI-adapted QUADAS framework showed that 67 of these 76 articles were excluded because they presented a high or unclear risk of bias in at least one evaluated domain. Consequently, only 9 studies were judged to have a low risk of bias and were included in the qualitative synthesis. The included evidence was dominated by studies using dynamic contrast-enhanced magnetic resonance imaging (DCE-MRI), with additional contributions from ultrasound-based approaches. The methodological strategies ranged from radiomics-based machine learning models to deep learning architectures.

Discussion

Integrating radiomic signatures with clinical predictors may improve performance in selected clinical scenarios, and peri-tumoral information alongside delayed DCE-MRI phases can contribute complementary diagnostic value. Nevertheless, the small number of low-risk studies highlights the need for standardized acquisition and preprocessing reporting, patient-level data splitting, robust external validation, and broader data availability to strengthen the clinical translation and generalizability of AI-based molecular subtyping.

Introduction

Breast cancer is the second most common cancer worldwide, accounting for 2.3 million new cases in 2022, and the fourth leading cause of cancer-related death, with approximately 670,000 fatalities, according to the International Agency for Research on Cancer (IARC). ¹ Biologically, breast cancer encompasses a heterogeneous group of diseases characterized by the uncontrolled proliferation of abnormal cells within breast tissue, leading to diverse molecular subtypes with variable prognostic and therapeutic implications.^2,3 The identification of these molecular subtypes has proven essential for achieving more accurate prognostic stratification and for guiding the selection of more effective targeted therapies.^3,4

To determine the molecular subtype of breast cancer, tissue samples obtained through biopsy are subjected to immunohistochemical (IHC) analysis. These assays evaluate the expression of key biomarkers, including estrogen receptor (ER), progesterone receptor (PR), and human epidermal growth factor receptor 2 (HER2), as well as the proliferative index Ki-67. ⁵ Based on these biomarkers, breast cancer is widely classified in the literature into four major molecular subtypes: luminal A (LA), luminal B (LB), HER2-positive (HER2+), and triple-negative breast cancer (TNBC). ³

However, biopsy is inherently an invasive procedure, as it requires the direct extraction of suspicious breast tissue. This intervention may lead to several adverse consequences for patients, ranging from subsequent clinical complications to psychological distress. Additionally, there is often a substantial delay between the biopsy procedure and the availability of IHC results, which can impact timely clinical decision-making. ⁶ Moreover, IHC presents diagnostic limitations due to its subjective nature: its interpretation relies heavily on the expertise and judgment of the pathologist, introducing variability and the potential for both intra-observer and inter-observer inconsistencies. ⁷ Therefore, there is a need for a strategy capable of capturing tumor molecular characteristics without relying on invasive procedures, with the aim of accurately classifying the four molecular subtypes of breast cancer.

A promising strategy arises from the analysis of medical imaging using artificial intelligence techniques. These approaches encompass imaging modalities such as full-field digital mammography, ultrasound (US), contrast-enhanced spectral mammography (CESM), and dynamic contrast-enhanced magnetic resonance imaging (DCE-MRI). Artificial intelligence techniques aim to characterize tumors identified in these medical images and classify them into the four molecular subtypes of breast cancer, thereby providing a noninvasive and potentially more efficient pathway for molecular subtype determination.

Despite the growing number of publications introducing artificial intelligence models aimed at identifying the molecular subtype of breast cancer from medical images, a significant gap remains in the translation of these technologies into clinical practice. This is primarily due to the lack of methodological rigor, limited external validation, poor generalizability stemming from homogeneous study populations, and the absence of standardized outcome measures. Together, these factors constrain clinicians’ trust and hinder regulatory acceptance.

Several systematic reviews and meta-analyses have examined the use of artificial intelligence in breast cancer imaging, including applications related to detection, diagnosis, prognosis, and treatment response. A smaller number of reviews has specifically addressed molecular subtype prediction from radiological images. However, these previous works have generally focused on a single imaging modality or have approached molecular markers within a broader radiogenomics framework. In addition, formal study-level risk-of-bias and applicability assessments have not been consistently incorporated, which limits the ability to distinguish findings with potential clinical relevance from results.

The present review addresses this gap by synthesizing AI-based approaches for breast cancer molecular subtype prediction across medical imaging modalities and by evaluating the methodological quality and applicability of the available evidence. Beyond model performance, it also critically examines methodological validity, applicability, and potential sources of bias that may affect clinical translation. For doing so, we conducted a systematic literature review following the PRISMA guidelines, ⁸ aiming to identify artificial intelligence methods developed for the classification of breast cancer molecular subtypes using radiological imaging. In addition, the risk of bias and applicability of the included studies were evaluated through a structured AI-adapted QUADAS- 2 framework. ⁹ By integrating technical performance analysis with methodological quality appraisal, this review provides a comprehensive and critical evaluation of the current evidence, highlighting not only promising AI-based approaches but also limitations related to validity, robustness, transparency, and clinical applicability.

Methods

Literature Search

This review was conducted following the methodological framework established by the Preferred Reporting Items for Systematic Reviews and Meta-Analyses (PRISMA) 2020 guidelines. ⁸ As a systematic review based solely on previously published studies, this study does not require an ethics statement. Also the systematic review is registered at the International Platform for Registration of Systematic Review and Meta-analysis Protocols (registration number: INPLASY202650096).

The systematic review focused on original, peer-reviewed research articles published between January 1, 2020, and March 25, 2025, that applied artificial intelligence techniques to classify the molecular subtype of breast cancer using medical imaging. The literature search was conducted across the scientific databases Scopus, Web of Science, ScienceDirect, and PubMed to ensure comprehensive coverage of relevant studies. Search terms incorporated combinations of keywords and Boolean operators, including: (”breast cancer”) AND (”subtype”) AND (”image” OR ”MRI” OR ”ultrasound” OR ”mammography” OR ”CESM”) AND (”prediction” OR ”classification” OR ”analysis”) AND (”radiomics” OR ”machine learning” OR ”deep learning”).

Study Selection

The inclusion criteria were as follows: (a) studies focused on molecular subtype classification; (b) studies using artificial intelligence algorithms; (c) studies based on in vivo medical images; (d) original research articles; (e) articles published in peer-reviewed journals; (f) articles available in full text; and (g) articles written in English.

Accordingly, the exclusion criteria were as follows: (a) studies that did not classify at least one molecular subtype, including studies focused exclusively on outcomes such as neoadjuvant chemotherapy response, pathological complete response, or disease-free survival; (b) studies based solely on conventional statistical analysis without artificial intelligence methods; (c) studies using non-in vivo images, such as histological or other ex vivo images; (d) non-original articles, including systematic reviews, narrative reviews, meta-analyses, editorials, letters, and commentaries; (e) publications not appearing in peer-reviewed journals, such as conference abstracts or proceedings; (f) articles not available in full text; and (g) articles not written in English.

Two reviewers independently performed the initial selection of articles by screening titles and abstracts using the Rayyan web-based platform for systematic reviews. Potentially eligible articles were then retrieved in full text for detailed assessment of eligibility and risk of bias with the QUADAS-2 AI modified tool. Any discrepancies between reviewers at any stage were resolved through discussion and consensus.

Data Extraction

For each included study, the following variables were extracted: publication year, study design (prospective or retrospective), sample size and class distribution, data source (public or private), imaging modality, molecular subtype classification task, AI model architecture, and reported performance metrics including area under the receiver operating characteristic curve (AUC), accuracy, sensitivity, and specificity. The AI model architecture was further categorized as either a machine learning or deep learning approach, consistent with the organizational framework used in the Results and Discussion section. For studies reporting multiple classification tasks or experimental configurations, performance metrics were extracted for the best-performing model or configuration per task.

Quality Assessment and Risk of Bias

Assessment of eligible studies was performed using the Quality Assessment of Diagnostic Accuracy Studies-2 (QUADAS-2) tool, ⁹ which evaluates four domains: patient selection, index test, reference standard, and flow and timing. Each domain is assessed for risk of bias, and the first three domains are additionally assessed for applicability concerns. For the specific purposes of this review, the original QUADAS-2 signaling questions were adapted to better capture methodological aspects that are critical in AI-based diagnostic studies. In the patient selection domain, questions were added to assess whether the imaging acquisition protocol and scanner model were reported; whether preprocessing strategies were described and unlikely to artificially increase model performance; whether the rationale and breakdown of training, validation, and test sets were explicitly stated; whether data partitioning was performed at the patient level rather than at the image or lesion level to reduce the risk of data leakage between model development and evaluation; whether class distribution was stratified across data splits; whether demographic or population characteristics were reported; whether data augmentation strategies, when applied, adequately represented expected clinical variability; and whether the dataset was publicly available or made accessible by the authors. In the index test domain, questions were added to evaluate whether external validation on an independent dataset was performed; whether hyperparameter tuning was adequately described; whether the reported evaluation metrics were appropriate for the classification objective, particularly in the presence of class imbalance; and whether statistical significance testing was conducted. Items from the reference standard and flow and timing domains were retained from the original QUADAS-2, with minor wording adjustments to reflect the AI context, including verification that molecular subtype labels were established independently of the AI model outputs. The complete set of adapted signaling questions is presented in Table 1.

Table 1.

QUADAS-AI Based Criteria for Risk of Bias and Applicability Assessment

Criteria	Answer
Domain 1: Patient selection
Are the eligibility criteria (patients and studies) to be included in the study clearly described?	Yes/No/Unclear
If preprocessing strategies are implemented, do they skew the model’s performance?	Yes/No/Unclear
Was the rationale and breakdown of training, validation, and test sets presented?	Yes/No/Unclear
Is training, validation, and testing data distributed?	Yes/No/Unclear
Was the data distribution stratified with respect to the class distribution?	Yes/No/Unclear
Is patient-based data distribution performed?	Yes/No/Unclear
Was the distribution of data stratified with respect to population data related to the study?	Yes/No/Unclear
If a data augmentation strategy is implemented, does it model the expected variability and is it distributed stratifically across the datasets (training, validation)?	Yes/No/Unclear
Was the data derived from open-source datasets or is it made public based on the study?	Yes/No/Unclear
Could the selection of patients have introduced bias?	RISK: Low/High/Unclear
Domain 2: Index test
Was external verification performed?	Yes/No/Unclear
Were the index test results generated without knowledge (information) of the results of the reference standard?	Yes/No/Unclear
Is an adequate or optimized adjustment of the hyperparameters of the proposed model performed?	Yes/No/Unclear
Do the evaluation metrics correspond to the objective being evaluated?	Yes/No/Unclear
Adequate statistical validity test?	Yes/No/Unclear
Could the conductor interpretation of the index test have introduced bias?	RISK: Low/High/Unclear
Domain 3: Reference standard
Was the reference standard likely to correctly classify the target condition?	Yes/No/Unclear
Were the reference standard results interpreted without knowledge of the results of the index test?	Yes/No/Unclear
Is there independence between the authors and the staff who develop the gold standard?	Yes/No/Unclear
Could the reference standard, your conduct, or your interpretation have introduced bias?	RISK: Low/High/Unclear
Domain 4: Flow and timing
Was the time between the index test and the reference standard reasonable?	Yes/No/Unclear
Could the patient flow and timing have introduced bias?	RISK: Low/High/Unclear

Results

Figure 1 shows the PRISMA flow diagram summarizing the identification, screening, eligibility assessment, and inclusion of studies in this systematic review on breast cancer molecular subtype prediction. The initial database search retrieved 626 records. After removal of 253 duplicates, 373 unique records remained. Screening of titles and abstracts led to the exclusion of 279 records that were not related to the prediction or classification of breast cancer molecular subtypes, leaving 94 records for full-text retrieval. For 18 of these records, the full text could not be obtained, so.

Figure 1.

PRISMA 2020 flow diagram for study identification, screening, eligibility assessment, and inclusion. A total of 626 records were identified across four databases (Scopus, Web of Science, ScienceDirect, and PubMed). After removal of 253 duplicates, 373 unique records were screened by title and abstract, of which 279 were excluded as not related to molecular subtype prediction. Of the 94 records retrieved for full-text assessment, 18 could not be obtained. The remaining 76 full-text articles were assessed for eligibility using the adapted QUADAS-AI framework; 67 were excluded due to high or unclear risk of bias in at least one domain. Nine studies met all inclusion criteria and were included in the qualitative synthesis

76 reports were assessed for eligibility. Following full-text review and risk-of-bias assessment using our modified QUADAS-2 tool, 67 studies were excluded because they presented a high or unclear risk of bias in at least one evaluated domain (Table 2). Consequently, 9 studies met all inclusion criteria and were finally included in the qualitative synthesis. The main characteristics of these studies are summarized in Table 3. Any missing or unclear data was indicated with a dash.

Table 2.

QUADAS-AI Risk of Bias Assessment for the 76 Full-Text Articles Evaluated. Each Domain is Rated as Low Risk (LR), High Risk (HR), or Unclear Risk (UR) of Bias

Study	Patient selection	Index test	Reference standard	Flow and timing
10	HR	UR	LR	LR
11	HR	UR	LR	LR
12	HR	LR	LR	LR
13	HR	UR	LR	LR
14	HR	UR	LR	LR
15	HR	UR	LR	LR
16	HR	UR	LR	LR
17	HR	UR	LR	LR
18	LR	LR	LR	LR
19	UR	UR	UR	LR
20	UR	LR	UR	UR
21	UR	LR	UR	UR
22	HR	UR	UR	UR
23	UR	HR	LR	LR
24	UR	LR	UR	LR
25	UR	LR	UR	UR
26	LR	LR	LR	LR
27	UR	UR	UR	UR
28	LR	UR	UR	UR
29	LR	LR	LR	LR
30	HR	UR	LR	LR
31	HR	UR	LR	LR
32	UR	UR	UR	LR
33	HR	LR	UR	LR
34	HR	LR	UR	UR
35	HR	LR	UR	LR
36	LR	LR	UR	LR
37	HR	HR	HR	HR
38	HR	HR	UR	LR
39	UR	LR	UR	LR
40	UR	UR	LR	LR
41	LR	LR	LR	LR
42	LR	UR	UR	UR
43	HR	UR	UR	UR
44	HR	UR	LR	UR
45	UR	UR	UR	UR
46	UR	HR	UR	UR
47	HR	HR	UR	UR
48	UR	UR	LR	LR
49	LR	LR	LR	LR
50	UR	HR	UR	LR
51	LR	LR	LR	LR
52	UR	LR	LR	LR
53	HR	UR	UR	UR
54	HR	LR	UR	UR
55	UR	UR	UR	UR
56	UR	LR	UR	UR
57	LR	LR	HR	UR
58	HR	LR	LR	LR
59	UR	LR	HR	LR
60	LR	LR	LR	LR
61	HR	HR	UR	LR
62	HR	HR	UR	LR
63	HR	UR	UR	UR
64	UR	UR	UR	UR
65	LR	LR	LR	LR
66	UR	UR	UR	LR
67	HR	HR	UR	LR
68	UR	LR	UR	LR
69	HR	HR	UR	LR
70	HR	UR	UR	LR
71	UR	UR	LR	LR
72	UR	UR	HR	LR
73	UR	UR	UR	UR
74	UR	LR	UR	LR
75	HR	UR	LR	LR
76	UR	UR	UR	LR
77	HR	UR	UR	LR
78	UR	HR	LR	LR
79	UR	UR	UR	LR
80	LR	HR	UR	LR
81	UR	UR	UR	LR
82	UR	HR	HR	LR
83	HR	HR	HR	LR
84	LR	LR	LR	LR
85	HR	HR	HR	LR

Table 3.

Technical Specifications of Included Studies. Abbreviations: DWI, Diffusion-Weighted Imaging; ADC, Apparent Diffusion Coefficient; NME-DWI, Non-mono-exponential Diffusion-Weighted Imaging; AUC, Area Under the Receiver Operating Characteristic Curve; Accu, Accuracy; Sens, Sensitivity; Spec, Specificity; SVM, Support Vector Machine; CNN, Convolutional Neural Network; DNN, Deep Neural Network; K-NN, K-Nearest Neighbors; RF, Random Forest; CART, Classification and Regression Tree; CC, Craniocaudal; MLO, Mediolateral Oblique

Study	Year	Type of study	Sample size	Data source	Imaging method	Validation	Molecular subtype	Model	AUC	Accu	Sens	Spec
29	2024	Prospective single-center	475 patients/480 lesions (85 LA, 141 LB (HER2−), 92 LB (HER2+), 88 HER2-enriched, 74 TNBC)	Private	DCE-MRI + NME-DWI	No	LA, LB (HER2−), LB (HER2+), HER2-enriched, TNBC (5-way; one-vs-others)	DNN ensemble (MP-MRI)	0.83	0.76	0.81	0.86
18	2024	Retrospective multicenter	214 HER2-negative (112 HER2-low, 102 HER2-zero)	Private	DCE-MRI	Yes	HER2-negative (HER2-low vs HER2-zero)	Nomogram (K-NN)	0.859	--	0.733	0.905
26	2024	Retrospective single-center	377 (73 LA, 151 LB, 78 HER2+, 75 TNBC)	Private	DCE-MRI	No	LA, LB, HER2+, TNBC (Luminal vs others; HER2 vs others; TNBC vs others; LA vs LB)	Logistic regression	0.735	0.796	0.654	0.845
49	2023	Retrospective single-center	160 (84 luminal, 76 non-luminal)	Private	DCE-MRI	No	LA, LB, HER2+, TNBC (Luminal vs non-luminal; Luminal A vs others; TNBC vs others)	CAMBNET (cross-attention multi-branch CNN)	0.959	0.881	0.858	--
41	2023	Retrospective single-center	185 (augmented with 25 SMOTE; total 210) Train 150/Val 60	Private	DWI/ADC-MRI	No	ER/PR status (ER/PR+ vs ER/PR−)	ADC radiomics (logistic regression) + Ki67 & grade (combined model)	0.938	--	0.786	0.955
51	2023	Retrospective single-center	145 (23 LA, 90 LB, 16 HER2+, 16 TNBC)	Private	US (grayscale + color Doppler)	No	LA, LB, HER2+, TNBC (TNBC vs non-TNBC)	VGG-19 (CNN)	0.860	0.850	0.860	0.860
65	2022	Retrospective multi-center	422 (293 HR+, 59 HER2+, 70 TNBC)	Private	DCE-MRI	Yes	HR+, HER2+, TNBC (TNBC vs non-TNBC; HR+ vs non-HR+; HER2+ vs non-HER2+)	SVM	0.852	--	0.692	0.890
60	2022	Retrospective multi-center	182 (118 malignant, 64 benign)	Private	CESM	No	Tumor nature; Grading; HER2 status; HR status (HER2+ vs HER2-; HR+ vs HR-)	Radiomics + ML (logistic regression, CART, RF)	0.859	0.958	1.000	0.929
84	2020	Retrospective single-center	365 invasive breast cancer Training: 150 (50 TNBC, 50 HER2, 50 Luminal) Validation: 71 (12 TNBC, 9 HER2, 50 Luminal)	Private	Synthetic mammography	No	LA + LB, HER2+, TNBC (TNBC vs non-TNBC; HER2 vs non-HER2; Luminal vs non-luminal)	Elastic-net radiomics signature (CC+MLO)	0.838	0.803	0.833	0.797

Machine Learning

In the field of DCE-MRI-based radiomics, the studies by Zhang et al. ⁶⁵ and Huang et al. ²⁶ have focused on the classi-fication of breast cancer molecular subtypes using dynamic contrast-enhanced magnetic resonance imaging, although with complementary methodological strategies. Zhang et al evaluated the preoperative prediction of molecular subtypes in 422 patients using early-phase DCE-MRI. Based on three-dimensional intratumoral segmentations and four annular peritumoral regions of 2, 4, 6, and 8 mm, they extracted 1,316 radiomic features. The authors developed intratumoral, peri-tumoral, and combined radiomic models using support vec-tor machines with L1 regularization-based feature selection. In addition, they constructed clinical-radiological models and a combined clinical-radiological-radiomic model. Their findings showed that peritumoral information provides com-plementary value to intratumoral radiomics, with the optimal peritumoral ring size depending on the classification task: 6 mm for HR-positive versus others and TNBC versus others, and 8 mm for HER2-enriched versus others. In binary clas-sification tasks, the combined clinical-radiological-radiomic model achieved the best overall performance, whereas in the ternary classification task, the intra+peri radiomic model achieved the highest accuracy. ⁶⁵

Similarly, Huang et al analyzed 377 patients using intra-and peritumoral segmentations with a 5-mm peritumoral region across three DCE-MRI phases: early, peak, and late. From each region and phase, 106 radiomic features were extracted. The authors trained subtype-specific logistic regression models and compared ten configurations for each differentiation task, including one-to-one models by phase and region, combined intra+peri models by phase, and a total fusion model. Although the differences among DCE-MRI phases were not statistically significant, models based on the late phase, particularly D _Intra and D_Intra+Peri, tended to show the best performance. In the full fusion model, the most important predictors according to SHAP were mainly derived from the late phase, reinforcing the conclusion that this phase should not be omitted. ²⁶ Taken together, these studies support the relevance of both intratumoral and peritumoral radiomics in DCE-MRI-based molecular subtype classification, while also suggesting that the temporal phase of contrast enhancement may influence radiomic model performance.

Complementarily, diffusion-derived radiomics has also shown potential for receptor-level stratification. Szep et al. ⁴¹ included 185 patients, with the dataset augmented to 210 cases using SMOTE, and performed whole-volume tumor delineation on apparent diffusion coefficient (ADC) maps to extract first-order radiomic features. The ADC-based model achieved an AUC of 0.81 in the training cohort and 0.93 in the validation cohort for differentiating ER/PR-positive from ER/PR-negative tumors. Moreover, the integration of radiomic features with Ki-67 and histological grade yielded an AUC of 0.93, supporting the added value of combining imaging-derived biomarkers with pathological variables for hormonal receptor status prediction.

Within a more specific molecular subgroup, Hu et al. ¹⁸ investigated the stratification of HER2-low and HER2-zero tumors among patients with HER2-negative breast cancer. Their study included 214 patients and used manually segmented lesions from the second phase of DCE-MRI and diffusion-weighted imaging (DWI). From these images, 107 radiomic features were extracted, of which six were retained to construct a radiomic score using a K-nearest neighbors classifier. In parallel, maximum tumor diameter and CA153 were identified as independent clinical predictors through multivariate logistic regression. The best overall performance was achieved by the nomogram integrating the radiomic score and clinical variables, highlighting the potential benefit of combining quantitative imaging features with clinical information for refined HER2-negative subgroup characterization. ¹⁸

Beyond MRI, X-ray-based radiomics has also been explored for subtype-related endpoints. Son et al. ⁸⁴ investigated synthetic mammography reconstructed from digital breast tomosynthesis in 365 patients with invasive breast cancer, grouped into three molecular categories: luminal A/luminal B, HER2-positive, and TNBC. From the synthetic mammograms, 129 radiomic features were extracted, and an elastic-net radiomics signature was developed. Clinical variables included age, lesion size, and radiologist-assessed imaging features. In the temporally independent validation cohort, the radiomics signature achieved its highest discriminative ability for TNBC, with an AUC of 0.838, whereas performance was more modest for HER2-positive tumors and luminal subtypes, with AUCs of 0.556 and 0.645, respectively. Importantly, the radiomics signature emerged as the only independent predictor, and its combination with clinical features improved TNBC discrimination compared with clinical variables alone. ⁸⁴

In a related contrast-enhanced mammography setting, Petrillo et al analyzed 182 patients, including 118 malignant and 64 benign lesions. Regions of interest were manually segmented on craniocaudal (CC) and mediolateral oblique (MLO) views, from which 837 textural features were extracted. Using Wilcoxon analyses and multivariate models, including logistic regression and tree-based algorithms with a balancing strategy, the authors reported strong performance for malignancy discrimination and clinically relevant histological endpoints. The best multivariate results included a maximum test accuracy of 95.83% for malignant versus benign lesion classification on CC images using non-regularized logistic regression, and 91.67% test accuracy for grading prediction on MLO images using a classification tree. Prediction of hormone receptor presence and HER2 status also achieved competitive accuracies, supporting the feasibility of contrast-enhanced mammography-derived radiomics for the noninvasive characterization of receptor-related outcomes. ⁶⁰

Deep Learning

Within MRI-based deep learning (DL), Ba et al. ²⁹ prospectively evaluated a deep neural network strategy combining DCE-MRI and nonmono-exponential model-based DWI (NME-DWI) in 475 patients with 480 lesions for five-way molecular subtype classification, including luminal A, luminal B/HER2-negative, luminal B/HER2-positive, HER2-enriched, and TNBC. Using an independent test split and an ensemble architecture that integrates modality-specific networks at the global feature level, the multiparametric MRI model achieved higher testing accuracy than either modality alone. Specifically, the combined MP-MRI model reached an accuracy of 0.76, compared with 0.71 for DCE-MRI and 0.64 for NME-DWI.

These findings suggest that DCE-MRI provides stronger subtype-related information than NME-DWI, while also demonstrating that multimodal fusion can further improve molecular subtype classification performance. ²⁹

In contrast to classical radiomics approaches, Sun et al. ⁴⁹ proposed CAMBNET, a DL model applied to DCE-MRI data from 160 patients with breast cancer using manually segmented tumor contours. From each tumor contour, the authors extracted the minimum bounding box containing the lesion, expanded the margin by 10 pixels in each direction, resized the resulting region to 64×64 pixels, and normalized the image before model input. CAMBNET was designed as a three-branch architecture composed of one surface feature extraction pathway and two tumor localization pathways. These branches were integrated through spatial and channel attention modules, together with a deep feature extraction module, enabling the network to combine tumor appearance, spatial localization, and discriminative feature representations. The authors also evaluated model robustness in a multi-source dataset including both 3T and 1.5T images, and explored the influence of age at menarche and tumor size on diagnostic performance. Specifically, they compared patients with menarche before versus after 14 years of age, as well as small tumors measuring < 20 mm versus large tumors measuring ≥ 20 mm, for the differentiation of luminal versus non-luminal and TNBC versus non-TNBC tumors. The model maintained acceptable performance on multi-source data and showed particularly favorable results in patients with early menarche and small tumors. Grad-CAM analyses further supported the interpretability of the model by showing that the network focused its attention on relevant tumor regions. ⁴⁹

Beyond MRI, DL has also been explored for molec-ular subtype-related prediction using ultrasound imaging. Boulenger et al. ⁵¹ developed a DL system for identifying TNBC from ultrasound images in 145 patients, includ-ing both grayscale and color Doppler acquisitions, without requiring manual tumor segmentation. The images were converted to grayscale, cropped, resized to 224×244 pix-els, and processed using adaptive histogram equalization to reduce intensity heterogeneity across imaging devices. A convolutional neural network based on a modified VGG-19 architecture was then trained from scratch. To improve interpretability, the authors analyzed the learned internal representations using t-SNE projections from the last hidden layer and generated saliency maps. These analyses showed that the network mainly focused on hypoechoic tumor tissue and tumor margins, suggesting that the model learned bio-logically and visually plausible imaging patterns associated with TNBC identification. ⁵¹

Discussion

This systematic review summarizes the current evidence on artificial intelligence methods developed to predict breast cancer molecular subtypes from medical imaging. Although the initial search and selection process identified 76 studies, only nine met the low-risk-of-bias criteria after full-text assessment using the adapted QUADAS-2 framework. This finding indicates that the field has grown rapidly in terms of model development, but the methodological quality required to be considered for clinical translation remains uneven. Among the studies finally included, DCE-MRI was the dominant modality, followed by ultrasound and contrast-enhanced mammography. The reported models, including radiomics-based machine learning approaches and deep learning architectures, showed promising performance for selected tasks such as TNBC identification, luminal versus non-luminal classification, HER2-related stratification, and receptor-level prediction.

Two methodological trends emerged in the included studies. First, models integrating radiomic features with clinical or pathological variables appeared to improve performance compared with image-derived features alone. Second, information from peritumoral tissue patterns and enhancement behavior in the late phases of dynamic contrast-enhanced magnetic resonance imaging (DCE-MRI) seems to provide complementary information related to subtype, suggesting that molecular subtype from imaging is not only dependent on the tumor features.

Previous reviews have addressed the use of radiomics, radiogenomics, and AI-based methods for breast cancer molecular characterization, although with different scopes and levels of assessment.^86-89 Darvish et al. ⁸⁹ reviewed the broader field of breast cancer radiogenomics, focusing on associations between imaging phenotypes and genomic or molecular indicators, including ER, PR, HER2, Ki-67, and gene-expression markers. However, that review was not specifically centered on AI-based molecular subtype classification. Nevertheless, this study reproted DCE-MRI as the highest-performing modality with AUCs up to 0.956 for specific genomic targets, consistent with the dominance of DCE-MRI in the evidence included in this work. Other recent reviews have provided modality-specific evidence. Thus, Fu et al. ⁸⁶ evaluated ultrasound-based AI for predicting molecular markers such as HER2, Ki67, PR, and ER; whereas Zhou et al. ⁸⁷ focused on ultrasound radiomics for molecular subtype prediction and reported pooled diagnostic performance for TNBC and luminal subtypes. These studies support the potential of ultrasound-based approaches. Mota ⁸⁸ reviewed AI-based characterization from mammography and tomosynthesis, including subtyping, staging, and prognosis, offering a detailed synthesis of these front-line imaging modalities but not a cross-modality assessment of AI-based molecular subtype prediction. These reviews further emphasize the translational promise of specific modalities, but they do not provide a unified evaluation of the risk of bias across the broader imaging evidence. In this context, the contribution of the present review is the application of a stricter methodological filter to identify recent AI-based studies on molecular subtype prediction from in vivo medical images. By adapting QUADAS-2 to include issues such as patient-level data splitting, data leakage, external validation, reporting of acquisition and preprocessing, class imbalance, and adequacy of performance metrics, this review distinguishes studies reporting promising results from those providing more reliable evidence for clinical translation.

The identification of studies with low risk of bias is relevant because these studies provide more reliable evidence on the performance and applicability of AI-based models. In this field, low-risk studies are those that more clearly report patient selection, image acquisition, preprocessing, data splitting, reference standard, and validation procedures. These elements reduce the possibility of data leakage, overestimated diagnostic performance, and limited reproducibility. Therefore, they can better support future clinical traslation. They are also important for regulatory evaluation, since they provide stronger evidence about the intended use of the model, its external validity, and its potential clinical benefit. Nevertheless, the limited number of studies classified as low risk of bias in this review indicates that further multicenter studies, preferably with external and prospective validation, are still needed before these models can be incorporated into routine clinical practice.

Limitations

This systematic review has several limitations. First, despite searching four major academic databases, the possibility of publication bias cannot be excluded, as studies with negative or inconclusive results are less likely to be published. Second, the search was restricted to English-language articles, which may have led to the exclusion of relevant work published in other languages. Third, the heterogeneity across included studies in terms of imaging modalities, acquisition protocols, classification tasks, and reported metrics precluded quantitative pooling of results through meta-analysis. Fourth, the QUADAS-AI framework applied in this review represents an adaptation of the validated QUADAS-2 tool developed by the authors for the specific context of AI-based diagnostic studies, and has not undergone external validation. Finally, the strict inclusion criteria requiring low risk of bias across all four domains resulted in a small number of included studies (n = 9), which limits the breadth of conclusions that can be drawn.

Conclusions

In conclusion, this systematic review underscores the grow-ing potential of artificial intelligence to non-invasively sup-port the prediction of breast cancer molecular subtypes from medical images. Across the included studies, radiomics-based models (predominantly using SVM and logistic regres-sion) and deep learning architectures achieved promising diagnostic performance for tasks such as differentiating luminal vs. non-luminal tumors, stratifying HER2-low vs. HER2-zero disease, and identifying TNBC. Multimodal and multiparametric strategies tend to improve discrimination and enhance the potential clinical relevance of the mod-els, while peri-tumoral information and delayed DCE-MRI phases emerge as complementary sources of subtype-related signal rather than redundant additions. Nevertheless, the small number of studies with low risk of bias, limited sample sizes, predominantly single-center designs, scarce external validation and calibration analyses, and persistent hetero-geneity in imaging acquisition, preprocessing, segmentation, feature extraction, and data partitioning continue to hinder clinical translation. Future research should prioritize rigor-ous methodological standardization and transparent report-ing, patient-level and multicenter validation, evaluation of calibration and clinical utility, and the construction of larger, more diverse and shareable datasets to fully leverage the potential of AI for the personalized molecular characteriza-tion of breast cancer.

Appendix.

Abbreviations

IARC

International Agency for Research on Cancer

IHC

Immunohistochemical

ER

Estrogen Receptor

PR

Progesterone Receptor

HER2

Human Epidermal Growth factor Receptor 2

LA

Luminal A

LB

Luminal B

HER2+

HER2-positive

TNBC

Triple-negative

CESM

Contrast-Enhanced Spectral Mammography

DCE-MRI

Dynamic Contrast-Enhanced Magnetic Resonance Imaging

US

Ultrasound

DL

Deep Learning

CC

Craniocaudal

MLO

Mediolateral Oblique.

ORCID iDs

Juan Camilo Morales-Duran https://orcid.org/0009-0006-8672-9557

Kevin Osorno-Castillo https://orcid.org/0000-0003-2132-7590

Gloria M. Díaz https://orcid.org/0000-0003-1028-9111

Ethical Considerations

The authors have nothing to report.

Data Availability Statement

The data that support the findings of this study are available from the corresponding author upon reasonable request.*