Breast cancer ranks among the most common cancers impacting women globally (1). It is a diverse disease characterized by a range of histopathological traits, molecular classifications, and clinical patterns, necessitating ongoing research to enhance diagnostic accuracy, prognostic assessments, and therapeutic approaches (2, 3). Immunohistochemistry (IHC) plays a vital role in the characterization of breast cancer, utilizing biomarkers such as hormone receptors (HR), human epidermal growth factor receptor 2 (HER2), Ki-67, and lymph node metastasis (LNM) to identify subtypes, predict prognosis, and assess treatment responsiveness (4). HR-positive cancers respond well to hormonal therapies, while HER2-positive cancers are more aggressive but can be targeted by drugs like trastuzumab. Ki-67 levels indicate tumor proliferation and LNM is crucial for staging and prognosis, influencing treatment strategies (5–8).

Magnetic resonance imaging (MRI) significantly improves the diagnosis, staging, and management of breast cancer by providing high-quality soft tissue contrast along with comprehensive functional insights (9). Intravoxel incoherent motion (IVIM) can obtain multiple quantitative parameters, such as the apparent pure diffusion coefficient (D), pseudo-diffusion coefficient (D*), and perfusion fraction (f) from multiple diffusion weighted imaging (DWI) images with different b-values using a biexponential fitting algorithm (10–12). These parameters provide non-invasively information about diffusion and perfusion simutaneously, allowing for a better characterization of breast cancer.

Radiomics extracts quantitative features from medical images, offering insights beyond what is visible to the human eye (13). MRI-based radiomics aids in tumor characterization and predicting treatment response (14). However, it is often challenging to relate radiomics features to biological or physiological meanings. A rule of thumb for radiomics study is that radiomics features should be extracted from a well-mixed region, but lesions of tumor may comprise of mesoscopically inhomogeneous regions of different characteristics, limiting its potential for medical research and treatment decision-making (14, 15).

Habitat analysis in radiomics addresses tumor heterogeneity by analyzing different regions, or “habitats”, within the tumor (16). This method takes into account the spatial complexity of tumors and offers valuable insights into their biological characteristics (17). IVIM imaging provides valuable parameters that describe both perfusion and diffusion within each voxel, while diffusion indicates the integrity and density of the cellular structure (18). Consequently, subregions clustered by IVIM parameters correspond to regions with different physiological characteristics. For example, regions with high cell density and high perfusion may correspond to areas of active tumor growth, whereas regions with low cell density and low perfusion may correspond to less aggressive or necrotic tissue (19). At the same time, integrating habitat features with clinical factors may contribute to a more comprehensive understanding of the biology of diseases and promote the development of precision medicine (20, 21). This study aims to explore the value of IVIM-based habitat imaging in predicting IHC in patients with breast cancer.

In this study, we explored the features from the whole tumor, habitats region and clinical characters, and found that the hybrid model combining habitat features and CF performs best in the prediction of these four IHC.

Compared to whole-tumor analysis, more refined habitat VOIs can better reflect tumor heterogeneity. In our study, the tumor was subdivided into four subregions: Part-1 (low D, moderate f) likely represents regions of high cellular density with preserved perfusion. Part-2 (high D) suggests regions of reduced cellularity, where elevated water diffusion (high D) aligns with necrotic or edematous zones (27). Part-3 (high f) reflects hyperperfused subregions, indicative of active angiogenesis (28). These areas may correlate with highly vascularized tumor fronts or inflammatory microenvironments, often associated with rapid growth or immune infiltration. Part-4 (low D, low f) denotes densely packed, hypoxic niches with restricted diffusion and poor perfusion. Such habitats are histologically consistent with high-grade tumors, which drive metastasis and chemoresistance (29). By analyzing these subregions, our study provided a more in-depth understanding of tumor heterogeneity.

Our study revealed several distinct associations between IVIM-derived habitat features and breast cancer molecular subtypes. In the high-D and low-f subregion (Part-2), a higher 10th percentile value of D was associated with HER2-positive tumors. This suggests that in HER2-positive tumors, even the most diffusion-restricted voxels within this subregion exhibit relatively higher diffusivity. Such a pattern may reflect a structurally loose and infiltrative growth pattern characteristic of HER2-positive breast cancers, where even the densest parts of low-perfusion regions are less compact. This finding aligns with the known aggressive and spatially invasive behavior of HER2-positive tumors (30). Furthermore, in the same subregion (Part-2), a higher median value of D was associated with Ki-67 positivity. This may be related to the rapid proliferation of tumors with high Ki67 and the occurrence of necrosis. Such regions may correspond to loosely structured stromal areas or infiltrative tumor margins, where low cellular packing density allows for increased water mobility despite reduced vascularization. This finding may reflect a distinct tumor microenvironment in highly proliferative breast cancers, characterized by rapid cellular turnover in structurally less constrained regions (31, 32). We also found that the proportion of tumor volume characterized by both D and low f (Part-4) was significantly higher in HR-positive tumors. This may reflect the relatively slow-growing and fibrotic nature of HR-positive tumors, which tend to accumulate more structurally stable, hypoperfused, and diffusion-restricted regions (33). In contrast, HR-negative tumors, including triple-negative breast cancer (TNBC), are often more heterogeneous and rapidly proliferative, resulting in fragmented or necrotic cores with unstable D/f patterns (34, 35). In the low-D and low-f subregion (Part-4), both D entropy and f energy were found to be significantly associated with LNM status. A higher entropy of D reflects increased heterogeneity in water diffusion, potentially indicating a complex microenvironment comprising necrosis, inflammation, fibrosis, and heterogeneous cell density. Such microstructural complexity has been correlated with more aggressive tumor behavior and higher metastatic potential (36). Similarly, a higher energy of f in the same region suggests the presence of repetitive, spatially organized perfusion signals, possibly representing residual microcirculation or neovascular structures within necrotic or fibrotic areas (37, 38). The elevated energy of perfusion-related features in low-perfusion regions underscores the role of organized microvascular structures in supporting tumor progression and metastasis. In addition, among clinical features, rim enhancement showed a notable association with Ki-67 positivity and HR negativity. Rim enhancement is characterized by peripheral contrast uptake with central hypoenhancement on DCE-MRI, often reflecting a combination of central necrosis and peripheral angiogenesis (39). This enhancement pattern is more frequently observed in tumors with high Ki-67 expression, as rapid tumor proliferation tends to outpace central vascular supply, leading to necrotic cores surrounded by viable, actively growing tumor rims (40). Furthermore, rim enhancement is commonly associated with hormone receptor-negative breast cancers, especially TNBC (41).

It is worth noting that habitat models constructed in this study used only first-order histogram features, which greatly improved the interpretability of the model. Besides, first-order histogram features are influenced only by the intensity values of the voxels, leading to greater stability in the analysis (42). This inherent stability enhances the robustness and reproducibility of habitat models, making them less susceptible to noise and variations in image acquisition parameters (43). In contrast, traditional radiomics used a bunch of texture features, which are not only difficult to interpret, but also readily influenced by image preprocessing steps, including image discretization and re-segmentation. Furthermore, the hybrid model, constructed with clinical and habitat features, improved the performance without compromising the interpretability, thus could be more readily accepted by radiologists.

The clinical relevance of IVIM-based habitat imaging lies in its ability to non-invasively map tumor heterogeneity by identifying subregions with distinct biological characteristics. By clustering voxels based on D and f parameters, habitat imaging can differentiate areas of high cellularity, necrosis, or active angiogenesis within tumors. This spatial resolution offers a deeper understanding of tumor biology, which is critical for predicting immunohistochemical markers such as HER2, Ki-67, HR, and LNM. These insights could guide clinicians in tailoring treatment strategies, such as prioritizing HER2-targeted therapies for tumors with high-D habitats or intensifying surveillance for patients with low-D/low-f subregions suggestive of aggressive behavior. Integrating habitat imaging into clinical workflows could enhance diagnostic accuracy and therapeutic planning. For example, during preoperative MRI evaluations, habitat maps could help surgeons target biopsy sites to regions of high proliferative activity, improving diagnostic yield. In radiation oncology, identifying hypoxic or perfusion-deficient subregions might enable dose escalation to radioresistant zones. Additionally, habitat features could supplement existing BI-RADS criteria to refine risk stratification, potentially reducing unnecessary interventions. The hybrid model, which combines habitat features with conventional MRI characteristics, further improves predictive performance, offering radiologists a tool to automate IHC predictions during routine image interpretation. This could streamline decision-making in settings where rapid biomarker assessment is critical, such as neoadjuvant therapy monitoring. Despite these advantages, several barriers may hinder clinical adoption. Technical standardization remains a challenge, as variations in MRI protocols (e.g., b-value selection, IVIM fitting algorithms) across institutions could compromise reproducibility. The clinical integration of IVIM-based habitat imaging requires a structured, multi-phase approach to ensure feasibility and reliability. First, standardization of MRI acquisition protocols is critical. This involves consensus-driven guidelines for b-value selection, IVIM fitting algorithms, and segmentation methodologies. At the same time, it is important to develop user-friendly software that enables automated habitat clustering and feature extraction. Integrating such tools into existing PACS or AI platforms could help streamline clinical workflows. These tools should emphasize interpretability by offering radiologists intuitive visualizations, such as color-coded habitat maps overlaid on MRI images, along with quantitative summaries like subregion volumes or perfusion metrics to support clinical decision-making. As habitat imaging remains in a developmental phase, the external validity of IVIM-based habitat models continues to be uncertain, thereby impeding their integration into standard clinical workflows.

However, there are also several limitations in our study that need to be acknowledged. The relatively small sample size (n = 299), with only 210 cases used for training, raises concerns regarding potential overfitting. Although internal cross-validation was performed to mitigate this risk, such an approach may not fully reflect the model’s generalizability to unseen data. Future studies should incorporate larger, multicenter datasets and external validation cohorts to better evaluate the robustness and clinical applicability of the model across different populations and imaging conditions. Additionally, we did not explore the specifics of IVIM reconstruction parameters. For example, the selection of b-values and the approach used to fit diffusion and perfusion components. While intraclass correlation coefficients (ICCs >0.75) were used to ensure feature robustness, segmentation discrepancies between radiologists, particularly in heterogeneous tumors with ill-defined margins, could affect habitat clustering and feature extraction. Furthermore, large-scale pathological studies, including animal experiments, may be needed to reflect the global spatial consistency of tumors.

In conclusion, compared with traditional radiomics models, the model combining habitats and CF features can achieve better performance and better interpretability. It can potentially help to improve the image-based diagnosis, and help doctors formulate personalized treatment plans for breast cancer patients. Habitat radiomics based on IVIM can help to better understand the biology of tumors, providing more information and support for clinical decision-making. Future research will further improve and validate the models to further improve their applicability and effectiveness in clinical settings.