MRI-based texture analysis for breast cancer subtype classification in a multi-ethnic population

Abstract

Introduction

Breast cancer, the most prevalent cancer among women globally, is classified into molecular subtypes (luminal, HER2-enriched, and triple-negative) to guide treatment and prognosis. Traditional subtyping methods, such as gene profiling and immunohistochemistry, are invasive and limited by intratumoural heterogeneity. MRI radiomics analysis offers a non-invasive alternative by extracting quantitative imaging features, yet its application in diverse, multi-ethnic populations remains underexplored.

Objective

This study aimed to identify predictive radiomic features from multiple MRI sequences to classify breast cancer subtypes, compare the performance of four MRI sequences, and determine the optimal machine learning (ML) model for this task. A total of 162 retrospective breast cancer MRI cases were semi-automatically segmented, and 256 radiomic features were extracted. A multimodal ML framework integrating random forest and recursive feature elimination was developed to identify the most predictive features based on the area under the receiver operating characteristic curve (AUROC).

Results

Key predictive features included age, tumour size, margin characteristics, and intensity patterns within the tumour. Among MRI sequences, inversion recovery and T1 post-contrast performed best for subtyping. In addition, texture-based ML models effectively emulated visual assessment, demonstrating the potential of radiomics in non-invasive breast cancer subtyping. With the top ten features, the AUROC values are 0.735, 0.630, and 0.747 for luminal, HER2-enriched, and triple-negative, respectively.

Conclusion

These findings highlight the role of MRI-based texture features and advanced ML in enhancing breast cancer diagnosis, offering a non-invasive tool for personalised treatment planning while complementing existing clinical workflows.

Supplementary Information

The online version contains supplementary material available at 10.1007/s10334-025-01285-9.

Introduction

The most commonly diagnosed cancer among women worldwide is breast cancer [1]. Breast cancer is divided into subtypes; luminal A and B, HER2-enriched and triple-negative breast cancer (TNBC), which affect the management options and prognosis [2]. While gene profiling and immunohistochemistry surrogates are the current clinical standard in subtyping, they are not without limitations. Biopsy is an invasive procedure that samples only a limited region of the suspicious lesion, despite the presence of intratumoural heterogeneity [3]. Intratumoural heterogeneity and molecular changes post-treatment necessitate more dynamic and non-invasive approaches to subtype breast cancer [4, 5]. In cases of non-responsiveness to chemotherapy or concerns regarding inadequate sampling, a repeat biopsy should be performed.

Magnetic resonance imaging (MRI) is important in breast imaging, with multiple applications that complement mammography and ultrasound with much higher sensitivity in lesion detection. It is employed as a screening tool for individuals at high risk of breast cancer, instrumental in local staging, onco-surgical planning, and tumour characterisation [6]. Reasons for MRI’s superiority in breast imaging include its additional functional information on tumour neoangiogenesis and heterogeneity [7].

MRI allows a “non-invasive sampling” of breast lesions by radiomics analysis. Radiomics analysis is a computer method to characterise grey-level spatial distribution within an image, capturing patterns imperceivable or indistinguishable by human eyes. Previous studies have reported the application of machine learning in differentiating benign and malignant lesions [8–10], risk of recurrence [11], response to chemotherapy [12], and risk of metastasis [13] from MRI, specifically harnessing the contrast-enhanced images.

There were also publications on texture analysis in breast cancer subtyping; however, the studies were performed in a homogeneous monoethnic study population [14, 15]. Previous reports have highlighted the lack of ethnic diversity in breast cancer clinical trials [16], with reported racial differences in imaging findings of breast cancer [17]. While MRI offers a wealth of data, identifying the most predictive imaging features remains a significant challenge, necessitating the integration of advanced computational methods such as machine learning. In this study, we hypothesised that MRI-based texture features differ significantly across subtypes in a multi-ethnic population.

Methods

The research was retrospective, quantitative, and cross-sectional in design. This study was conducted at the University Malaya Medical Centre, which involved 162 cases. Invasive breast cancer patients who were treatment-naïve primary breast cancer with MRI performed between January 2018 and December 2022 were included. Exclusion criteria were MRI breast cases performed post-treatment, patients with recurrent, metastatic, or previously treated breast cancers, no enhancing mass lesion detected, or suboptimal image quality for interpretation. Demography data, including ethnicity, were collected from the medical records.

MRI breast examination and protocol

MRI breasts were performed in a 3.0 Tesla MAGNETOM System (Siemens Healthcare, Germany) or a 3.0 Tesla Signa® MRI Systems (General Electrics (GE) Healthcare, USA) with the patient in a prone position, using a dedicated breast coil. The intravenous contrast injection protocol was 20 ml of 0.5 mmol/mL gadoterate meglumine (Dotarem, Guerbet, China) via a power injector with a rate of 1 mL/s and a 20 mL saline flush. The post-contrast dynamic scan time was 1-min interval per phase. The scanning protocols were tabulated in Supplementary Tables 1 and 2. Figure 1 shows the flow chart for case selection.

Fig. 1.

Flow chart illustrates the process of case selection for the study, resulting in the inclusion of 162 cases for analysis

Histopathology data

The oestrogen receptor (ER), progesterone receptor (PR), and human epidermal growth factor receptor-2 (HER2) status were evaluated based on the American Society of Clinical Oncology/College of American Pathologists (ASCO/CAP) [18]; the breast cancer subtypes were defined based on immunohistochemistry (IHC) with or without in-situ hybridisation (ISH) [18]. The subtype categorisation is summarised in Table 1. Tumour grading was based on the Bloom–Richardson grading system. The data on receptor expression status (ER, PR, HER2) and tumour grade were obtained from the histopathology report of the core needle biopsy specimen.

Table 1.

Breast cancer subtype categorisation based on immunohistochemistry (IHC) and dual-colour dual Hapten hybridisation (D-DISH)

Subtype	Receptor status
Luminal	ER/PR + , HER2- and ER/PR + , HER2 +
HER2-enriched	ER/PR-, HER2 + *
Triple-negative breast cancer	ER/PR/HER2-

^*If equivocal HER2 results from IHC, D-DISH was used to determine the HER2 status, with > 2.0 taken as positive

Texture features extraction

The first- and second-grade radiomics features were extracted from the MRI images using local image feature extraction (LIFEx) software (version 7.0.11) [19]. LIFEx is an Image Biomarker Standardisation Initiative (IBSI)–compliant texture extraction software [20]. Sequences selected for feature extraction were T1W post-contrast in phase 2, T1W pre-contrast (plain), Inversion Recovery images (Short tau inversion recovery (STIR) or Turbo Inversion Recovery Magnitude (TIRM), and T2W images. The cases selected were downloaded from picture archiving and computer systems (PACS) in digital imaging and communication in medicine (DICOM) format and anonymised. A board-certified breast radiologist (NM) segmented the 3-dimensional volume of interest (VOI) using a semiautomatic threshold-based method with manual correction. Manual correction was performed when visually perceptible areas were excluded from the VOI or when incorrect regions were included. The VOI was reviewed by a senior supervising breast radiologist (MR). Training for image segmentation and feature extraction was available through the software resource webpage and undergone by the radiologists involved before the commencement of the study. The segmentation was performed on the T1W post-contrast image and the same VOI was mapped onto the other three sequences for feature extraction. The spatial resampling was standardised at 1 × 1 x 1 mm³ with intensity discretisation of 2 bins prior to feature extraction to ensure uniform spatial resolution across different MRI protocols and scanners. Tumour marker and/or biopsy track (if present) will be excluded from the VOI. Radiomics features were extracted from this VOI. The features extracted were tabulated in an Excel worksheet. The study workflow is summarised in Fig. 2.

Fig. 2.

Study workflow

Extracted radiomics features were harmonised using the Combating Batch Effects (ComBat) algorithm to handle the multi-protocol dataset [21]. We used an online application: ComBaTool (shinyapps.io) in this study. The following are two examples of texture features from the two different scanners after normalisation by ComBat (Fig. 3). This process aimed to ensure that the texture features extracted were more representative of the underlying tissue characteristics rather than being influenced by the differences in the MRI scanner characteristics.

Fig. 3.

Examples of texture features from the two different scanners after normalisation by ComBat

Data preprocessing

This study utilised a dataset comprising records from 162 women diagnosed with breast cancer, encompassing a total of 286 variables. Among these, 259 were continuous, 23 were nominal, and 3 were ordinal. A list of all features is provided in Supplementary Table 3.

Outliers were detected in 254 (98%) of the continuous variables. To mitigate their impact, capping was applied, replacing extreme values with the respective lower and upper bounds. In addition, four variables; lymphovascular infiltration (LVI), symmetricity, diffusion-weighted imaging (DWI), and apparent diffusion coefficient (ADC)—contained missing values, with ADC exhibiting the highest rate at 5.6%. To handle these missing values, the K-nearest neighbours (KNN) imputation method was applied, using K = 5, where missing values were estimated based on the five closest neighbours. The weight parameter was set to ‘uniform’ to ensure equal contribution from all selected neighbours.

Furthermore, normality testing using the Kolmogorov–Smirnov test (p < 0.05) confirmed that all continuous variables followed a non-normal distribution. Since ensuring feature comparability is essential for accurate parameter weight estimation during machine learning training and for facilitating convergence, standardisation was applied to all continuous variables. This was achieved using the Z-score normalisation method: Z = (X − μ)/σ.

Regarding the categorical variables, 12 original categories were transformed into 40 dummy variables. Following this, 12 variables were removed from the dataset as they were deemed non-informative for the study objectives. These included variables related to the year of diagnosis, research ID, surgery type, treatment type, and clinical outcome, as well as non-MRI features such as ER, PR, HER, LVI, nodal status, tumour grading, and tumour staging. As a result, the final dataset analysed consisted of 303 variables and 162 participants.

Data partitioning

The dataset was divided into a training set of 113 (70%) and a testing set of 49 cases (30%). Random stratification was applied to maintain a proportional representation of molecular subtypes across both subsets. Specifically, the training group comprised luminal (73%), HER2-enriched (12%), and TNBC (15%) subtypes, while the testing group included luminal (74%), HER2-enriched (12%), and TNBC (14%) subtypes. The data partitioning process is presented in Fig. 4.

Fig. 4.

Distribution of cases in training and testing partitions (1 = luminal, 3 = HER2-enriched, and 4 = TNBC). The bar charts illustrate the distribution of breast cancer subtypes (luminal, HER2-enriched, TNBC) in the training dataset (70% of total cases, n = 113) and the testing dataset (30% of total cases, n = 49). The proportions for each subtype were maintained using random stratification, ensuring consistency across the training and testing groups

Feature selection

The processed dataset was high dimensional, consisting of 303 variables. Reducing the number of features offers several benefits, including lowering computational costs during model training and testing, minimising the risk of overfitting, especially due to multicollinearity among variables and improving model interpretability and explainability.

To identify the most relevant predictive features for classifying breast cancer subtypes (luminal, HER2-enriched, or TNBC), we employed a multimodal hybrid feature selection approach. This method combined the random forest (RF) algorithm (an embedded technique) with recursive feature elimination (RFE) (a wrapper-based machine learning method) on the training data (n = 113). A detailed schematic of this hybrid feature selection process is provided in Fig. 5.

Fig. 5.

Hybrid workflow of feature selection using random forest and recursive feature elimination. In the first part (black text), RF ranked feature importance based on Gini impurity decrease, generating subsets of top-ranked features (e.g. top 10, 20, 30, 40, and 50). In the second part (blue text), each subset was iteratively evaluated using RFE. The iterative RFE process continued until no further AUROC increment was observed, resulting in the selection of the optimal feature subset associated with each breast cancer subtype

Initially, the RF algorithm was used to rank features based on their importance using fivefold cross-validation. The random forest model was configured with 100 decision trees, utilised Gini impurity to evaluate split quality, where a greater decrease in Gini impurity indicated higher feature importance, and trained each tree on a bootstrap sample of the dataset. The ranked feature importance results are presented in Fig. 6.

Fig. 6.

Ranking of features by Gini Impurity for a luminal, b HER2-enriched (HER2), and c triple-negative (TNBC)

Next, recursive feature elimination (RFE) was applied to refine the feature subsets generated by the Random Forest (RF) algorithm. RFE operates by iteratively removing the least important feature at each step. The process began with initialising a Logistic Regression model and employing fivefold cross-validation to assess predictive performance. At each iteration, the least significant feature was removed, and the model’s performance in classifying cancer subtypes was evaluated using the area under the receiver operating characteristic curve (AUROC). This iterative process continued until the optimal feature subset; determined by the highest average, AUROC across cross-validation folds was identified.

In addition, various feature subsets, ranging from the top 10 to the top 50 ranked by RF, were tested with the RFE algorithm. The results indicated that the top 10 features ranked by RF yielded the highest predictive performance (i.e. the highest AUROC) during RFE-based feature selection, as detailed in Supplementary Table 4. The final set of predictive variables selected by RFE consisted of 18 continuous variables and 1 categorical variable, as shown in Supplementary Table 5.

Given the potential impact of multicollinearity on model performance, such as inflated standard errors and inaccuracies in feature coefficient estimates, it was necessary to address correlations among the selected predictive features. Spearman's rank correlation coefficient test was performed to assess inter-variable correlations among the 18 continuous variables. The results, visualised in a heatmap (Fig. 7), identified variables with correlation coefficients exceeding 0.7 as highly correlated. To mitigate multicollinearity, the less important features (as determined by RF feature importance ranking) were excluded from model training.

Fig. 7.

Heatmap illustrating Spearman's ranked correlation coefficients among 18 predictive features with continuous data

Data reporting and visualisation

Categorical data were reported as frequency and percentage. All continuous were non-normal distributions. Thus, the descriptive data were reported as median with an interquartile range (IQR). The assessment of normality was conducted utilising the Kolmogorov–Smirnov test, using the scipy Python library (version: 1.11.1). The KNN imputation, feature scaling (standardisation), data partitioning, machine learning algorithms for feature selection (RF, RFE), as well as training and testing of prediction algorithms, were performed using the sklearn Python library (version 1.3.0). For visualisation purposes, matplotlib (version: 3.7.2) and seaborn (version: 0.12.2) served as the libraries to generate figures.

Ethical considerations

The study received approval from the Research Ethics Committee (2019822–7771). The requirement for informed consent was waived due to the retrospective nature of the study.

Results

There were 162 MRI breast examinations included. The median age of the study participants was 54.0 years (IQR 44.0 to 64.0). All included cases were invasive breast carcinoma (not otherwise specified). The median tumour size measured 2.55 cm, (IQR 1.78–3.50). There were 73.5% (n = 119) luminal, 11.7% (n = 19) HER2-enriched and 14.8% (n = 24) TNBC. Table 2 shows the study demography.

Table 2.

Demography of study cases

Description	n (%)
Ethnicity
Chinese	67 (41.4%)
Malay	64 (39.5%)
Indian	31 (19.1%)
Breast cancer subtype
Luminal	119 (73.5%)
HER2 enriched	19 (11.7%)
Triple-negative	24 (14.8%)
Tumour grade
1	16 (9.9%)
2	94 (58.0%)
3	52 (32.1%)

After preprocessing, the dataset comprised 303 independent variables, which were subjected to feature selection using a hybrid Random Forest (RF) and recursive feature elimination (RFE) machine learning approach to classify three breast cancer subtypes. Through fivefold cross-validation, RF ranked all features from 1st to 303rd based on Gini decrease for each cancer subtype. The analysis was then focussed on the top 10 features for each subtype, as this subset achieved the highest AUROC when modelled with RFE (Supplementary Table 4).

For the luminal subtype, RFE identified 10 key variables: age, tumour size, AUC of cumulative sum histogram, intensity 1st quartile, maximum intensity, discretised intensity 1st quartile, mean intensity, discretised mean intensity, and skewness of intensity distribution from the inversion recovery (IR) sequence. In addition, short run emphasis (GLRLM) from the T1W pre-contrast sequence was also selected as an important predictive feature (AUROC = 0.74). However, five IR sequence features—intensity 1st quartile, maximum intensity, discretised intensity 1st quartile, mean intensity, and discretised mean intensity—were highly correlated (Spearman r > 0.7) with AUC of cumulative sum histogram and were considered redundant.

For the HER2 subtype, RFE selected six features, achieving an AUROC of 0.63. These included tumour size, discretised intensity 3rd quartile, and high grey-level zone emphasis (GLZLM) from the IR sequence, as well as volume in voxels, sphericity, and discretised mean intensity from the T1W post-contrast sequence. However, high grey-level zone emphasis (GLZLM) was highly correlated with discretised intensity 3rd quartile and was thus considered redundant.

For the TNBC subtype, RFE identified five key features: circumscribed margin, discretised intensity skewness, intensity skewness, and compactness from the IR sequence, along with discretised mean intensity from the T1W post-contrast sequence. Intensity skewness was found to be highly correlated with discretised intensity skewness, making it redundant. These five features achieved an AUROC of 0.75 in distinguishing TNBC.

Following the removal of redundant features (correlation > 0.7), the final set of predictive variables included five for luminal, five for HER2-enriched, and four for TNBC (Table 3). Significant demographic and structural features identified in previous studies [22] were incorporated into the feature selection process. Notably, the predictive features were distinct for the luminal and HER2-enriched with TNBC subtypes, as they did not share common features (Fig. 8).

Table 3.

Machine learning-selected features for breast cancer subtypes

	Luminal	HER2-enriched	TNBC
Clinical feature	Age
Structural feature	Tumour_size	Tumour_size
Shape feature		Volume in Voxelsb Sphericityb	Circumscribed margin compactnessa
Intensity feature	AUC of cumulative sum histograma Skewness of intensity distributiona	Discretised Intensity 3rd Quartilea Discretised Mean Intensityc	Discretised intensity skewnessa Discretised mean intensitya
Second-order statistics	Short run emphasis (GLRLM)c

^afrom T1 post-contrast sequence

^bfrom T1 pre-contrast sequence

^cfrom the inversion recovery sequence

Fig. 8.

Venn diagram of predictive features for breast cancer subtypes. The Venn diagram illustrates the overlap and distribution of predictive features for three breast cancer subtypes: luminal, HER2, and triple-negative (TN). Each circle represents the number of features associated with a specific subtype, with unique features shown in non-overlapping regions and shared features displayed in overlapping sections.

Tumour size (cm) was identified as a common predictive feature for the luminal and HER2-enriched subtypes. Among patients with the HER2-enriched subtype, the median tumour size was larger at 2.8 cm (IQR: 2.3–3.8 cm), whereas for the luminal subtype, the median tumour size was smaller at 2.4 cm (IQR: 1.7–3.4 cm). A comparative analysis between these subtypes demonstrated a moderate effect size (Cohen’s d = 0.46), suggesting a meaningful difference in tumour size distribution between the luminal and HER2-enriched (Fig. 9).

Fig. 9.

Violin plot comparing tumour size distribution between HER2-enriched and luminal breast cancer subtypes

Discussion

This study aimed to identify the most effective texture features from specific MRI sequences for classifying breast cancer molecular subtypes within a multi-ethnic population (comprising Chinese, Malay, and Indian groups) using a hybrid Random Forest–recursive feature elimination (RF-RFE) machine learning algorithm. Our findings indicate that the most discriminative features for breast cancer subtype classification, based on the AUROC metric, included tumour size, shape, margin characteristics, and intensity-related features, including the pattern of intensity distribution within the tumour. The MRI sequences that contributed the most significant features were inversion recovery and T1 post-contrast sequences.

Previous studies have employed RF models for various prediction and classification tasks in breast cancer, such as predicting pathological complete response (pCR) after chemotherapy [23], classifying prognostic biomarkers and subtypes [24], and identifying metastatic brain tumour types [25]. As a supervised machine learning algorithm, RF is well suited for both regression and classification tasks. It is capable of handling datasets with a mix of continuous and categorical variables, making it particularly effective for classifying breast cancer subtypes, as demonstrated in our study.

The hybrid RF-RFE approach is particularly advantageous in radiomics due to its ability to manage high-dimensional, complex, and noisy data. This method has also been successfully applied in previous studies, such as the prediction of pCR after chemotherapy using MRI, where RF and RFE demonstrated high accuracy in classifying pCR following neoadjuvant chemotherapy [26]. In addition, Mylona et al. evaluated multiple machine learning models and identified RFE as one of the most effective feature selection techniques, alongside L1-LASSO, RF importance (RF-imp), and Boruta [27]. Previous study by Mitani, Ono [28], reported that RF-RFE was superior to XGBoost when combined with RFE. Radiomic datasets, as used in this study, involve high-dimensional data with hundreds of features derived from MRI. The RF-RFE approach effectively reduced dimensionality by selecting only the most relevant features, thereby improving model performance and minimising overfitting. RF ranks feature importance using metrics, such as the Gini Impurity, while RFE iteratively removes less significant features to prevent correlated or redundant features from negatively impacting the model. RFE operates by recursively eliminating features and building models on subsets of the data to identify the most predictive feature set. By selecting a smaller, interpretable feature set through RFE, the resulting model becomes better aligned with clinical applications, enabling clinicians to understand and trust its predictions.

The sequence that yielded the greatest number of characteristics following selection and reduction in our investigation is inversion recovery (IR) (2 features for luminal, 1 feature for HER2-enriched, and 2 for TNBC). This is followed by the T1W post-contrast sequence, which provided two features for HER2-enriched and one for TNBC. The T1W pre-contrast sequence identified one feature for luminal and one for HER2-enriched. In summary, 7 out of the 10 radiomics features were derived from non-contrast-enhanced sequences. This highlights the relevance of non-contrast-enhanced sequences, specifically IR, in providing comparable radiomics data and independence from the reliability of angiogenesis or perfusion information in radiomics research. This applies specifically to some instances where the use of gadolinium is contraindicated or not advised [29, 30].

Morphological tumour features, including shape and margin, play a crucial role in lesion characterisation in MRI breast, as outlined in the standardised reporting guidelines of the ACR-BI-RADS committee [31]. The shape-related features selected in this study, which are sphericity, volume in voxels, and compactness, align with these established radiological criteria. Therefore, it is not surprising that morphology emerged as an important radiomic feature for classifying HER2-enriched and TNBC subtypes in our study. This finding is consistent with previous research, which also reported significant differences in shape features between breast cancer subtypes [32]. Figure 10 are sample cases from our study demonstrating the difference in features of each breast cancer subtype (Fig. 10).

Fig. 10.

Case example of MRI images (post-contrast) in different molecular subtypes from the study. A Luminal, B HER2-enriched and C triple-negative breast cancer

The differences in imaging and texture features observed among breast cancer subtypes in our study can be related to their known cellular and pathological characteristics. TNBC typically demonstrate high cellularity and central tumour necrosis, which may contribute to the rim-enhancement and increased texture heterogeneity seen on MRI. HER2-enriched tumours are associated with increased angiogenesis and higher cellular density, aligning with the observed higher enhancement patterns and certain texture parameters indicating heterogeneity. In contrast, luminal subtypes generally have lower cellular density and slower proliferation rates, which are reflected in less aggressive enhancement patterns and lower heterogeneity on texture analysis. By aligning these imaging findings with the established biological behaviour of the subtypes, our study supports the potential utility of MRI-based texture analysis as a surrogate marker for tumour biology.

The present study highlighted the significance of sequence selection and model optimisation in enhancing accuracy and likelihood ratios for breast cancer subtype classification. Previous research employing IR images from MRI in radiomics studies for breast cancer has focussed on predicting LVI [33] and predicting ALN in breast cancer [34]. Gamal et al. [35] demonstrated that radiomic features extracted from non-contrast sequences, including T1W, T2W, and IR, can predict neoadjuvant chemotherapy response in breast cancer.

The semantic visually perceptible tumour morphological features, including shape and margin, are critical findings for lesion characterisation in MRI breast according to the standardised reporting guidelines of the ACR-BI-RADS committee [31]. Consequently, it is not unexpected that a morphological feature (shape) is among the selected radiomic characteristics for the classification of HER2-enriched and TNBC subtypes in our study. The shape feature has also been previously documented to have a considerable variance among subtypes [32]. In this study, volume, compactness, and sphericity were selected as key features for the HER2-enriched and TNBC. This is likely because tumour morphology and spatial characteristics are particularly relevant for these aggressive subtypes, which exhibit distinct morphological and growth patterns compared to others [36]. Biologically, TNBCs are inherently more aggressive, making features such as compactness valuable in quantifying the shape. Compactness is a measure of how closely a tumour’s shape resembles a perfect sphere, distinguishing TNBC from the other subtypes [37]. Compactness measures surface irregularity, with lower compactness values indicating greater irregularity.

In contrast, sphericity evaluates the overall resemblance to a sphere, where lower sphericity values suggest elongation or asymmetry. Volume-in-voxel and sphericity were identified as key features in the HER2-enriched subtype. Tumour size, a distinguishing feature between HER2-enriched and luminal subtypes in this study, was incorporated into the first step of the nomogram series. Sphericity quantifies how spherical a VOI is, and as previously reported, the HER2-enriched subtype exhibits distinct shape characteristics compared to other subtypes [38].

Radiomic intensity features, such as discretised AUC CSH (uniformity across classes using compact summarised histograms), skewness, and mean intensity, have shown correlations with specific breast cancer molecular subtypes in this study.

In addition, our study highlights the importance of the spatial distribution of discretised grey levels within the tumour in distinguishing luminal subtype cancers from other subtypes. The grey-level run-length matrix (GLRLM), which quantifies the length of consecutive voxels with the same grey level along a defined direction, was identified as a key predictive feature. In GLRLM, fine textures exhibit shorter runs, whereas coarser textures have longer runs with varying intensity values. Prior studies have demonstrated the utility of run-length non-uniformity in differentiating benign from malignant breast lesions in dynamic contrast-enhanced MRI [39] and in distinguishing luminal A from luminal B subtypes [32]

This study presents a novel approach compared to existing literature, as research on predictive modelling based on multi-ethnic patient cohorts remains limited. We focussed on invasive breast cancer cases and utilised multiple MRI sequences for molecular subtyping, enhancing the model’s generalisability and applicability across diverse populations. Notably, our findings demonstrate that texture features extracted from non-contrast-enhanced MRI sequences can effectively differentiate subtypes, reinforcing the potential of alternative imaging techniques. By testing four different MRI sequences, we provide further evidence supporting this observation.

However, this study has several limitations. It was conducted at a single centre with a relatively small sample size and an imbalanced subtype distribution. In addition, luminal A and luminal B subtypes were grouped as a single luminal category due to the absence of a standardised separation method in our dataset. We acknowledge the potential bias introduced by VOI drawn from post-contrast images mapped to the rest of the sequences. The segmentation was performed by a single trained radiologist with review by a senior breast radiologist. A multi-reader segmentation with inter-observer variability analysis would enhance methodological rigour; this will be considered in future prospective studies and external validation. Other potentially relevant clinical variables, such as risk factors and menopausal status, were not collected. We acknowledge that precision and other threshold-dependent performance metrics were not reported in this study, as our primary objective was to evaluate the overall discriminative ability of MRI-based texture features using AUROC, which is threshold-independent and appropriate for multi-class classification. Future studies with larger sample sizes, diverse imaging datasets, and external validation are necessary to confirm the reproducibility and robustness of our findings and to further assess model performance in real-world settings.

Clinical implications and future direction

This study lays the groundwork for expanding the role of MRI breast in complementing or potentially replacing invasive techniques for molecular subtyping, thereby promoting less invasive and more efficient diagnostic strategies towards precision medicine. The study observed no significant differences in MRI-derived features among the three major ethnic groups, indicating that the imaging characteristics of breast cancer subtypes are largely consistent across ethnicities. Hence, our report supports the potential generalisability of imaging-based predictive models in multi-ethnic populations.

We acknowledge that the current study is preliminary, and for this method to be adopted clinically, further validation on a larger, multicenter dataset with diverse scanner settings is essential. While it is challenging to specify an exact number, we estimate that a dataset of at least 500–1000 patients with balanced subtype distribution, including external validation, would be necessary to refine the model, evaluate its generalisability, and establish robust performance metrics suitable for clinical implementation.

Future studies may benefit from exploring ensemble machine learning approaches that integrate multiple models, as well as implementing subtype-specific preprocessing pipelines. These strategies could further improve predictive accuracy and enhance the robustness of radiomics-based classification.

We envision that the proposed MRI-based texture analysis method could serve as a non-invasive adjunct tool in the clinical workflow to aid in the pre-treatment prediction of breast cancer subtypes. This approach may complement histopathological analysis, particularly in cases where biopsy results are inconclusive, repeated biopsies are not feasible, or when there is a need for additional non-invasive assessment for treatment planning and risk stratification. For clinical implementation, further validation on a larger, multicenter dataset is essential to ensure generalisability and robustness of the model across diverse patient populations and imaging protocols.

Conclusion

This study explored the potential of radiomics features extracted from various MRI sequences to classify breast cancer molecular subtypes. Key features such as morphology (shape and margin), tumour size, and intensity in IR and T1 post-contrast sequences were identified as valuable predictors. These findings highlight the promise of MRI sequence-specific features in advancing precision medicine and warrant further investigation.

Integrating radiomic texture analysis into clinical workflows could enable the prediction of molecular subtypes from imaging data obtained before histopathology and IHC results become available. In addition, this study lays the foundation for expanding the role of imaging in potentially replacing invasive procedures for molecular subtyping, paving the way for less invasive and more efficient diagnostic strategies in breast cancer management.

Supplementary Information

Below is the link to the electronic supplementary material.

Author contributions

All authors contributed to the study's conception and design. Material preparation, data collection and analysis were performed by Nazimah Ab Mumin, Chuin-Hen Liew and Song-Quan Ong. The first draft of the manuscript was written by Nazimah Ab Mumin and all authors commented on previous versions of the manuscript. All authors read and approved the final manuscript.

Funding

Open access funding provided by The Ministry of Higher Education Malaysia and Universiti Teknologi MARA. This work was supported by the Malaysian Ministry of Higher Education/ Ministry of Science, Technology and Innovation: FRGS/1/2019/SKK03/UM/01/1, which Kartini Rahmat received and UiTM grant: 600-RMC/LESTARI-SDG-T 5/3 (012/2120), which Nazimah Ab Mumin received.

Data availability

The datasets generated and analysed during this study are available from the corresponding author upon reasonable request, as consent for public sharing of the data was not granted from the research ethics committee.

DeclarationsDeclarations

Conflict of interest

All authors declare no conflict of interest.

Ethics approval and Consent to publish

This study was conducted following the principles outlined in the Declaration of Helsinki. The study received approval from the Research Ethics Committee (2019822-7771). The requirement for informed consent was waived due to the study’s retrospective nature.