Breast cancer has become the most common malignancy among women worldwide, with its incidence and mortality rates rising annually, posing a significant threat to women's health [1]. Early diagnosis and treatment of breast cancer are crucial to patient survival rates [2]. With the global advancement in precision medicine, biomarkers have become increasingly important in the individualized treatment of breast cancer. Ki-67 is a non-histone nuclear protein expressed during cell mitosis, related to the cell cycle, and often used as a marker of cell proliferation. It has been repeatedly proven to be an independent predictor and prognostic factor for breast cancer and other tumors [3]. The role of Ki-67 in the prognosis and prediction of breast cancer has been widely discussed [4]. The expression level of Ki-67 in breast cancer cells holds significant value in the diagnosis and prognosis of breast cancer [[5], [6], [7]]. Currently, Ki-67 detection is mainly performed through immunohistochemistry (IHC) scoring following surgical excision or biopsy [8]. Additionally, the proliferation status of Ki-67 varies in different regions of the tumor [9], and its invasive nature and sampling limitations restrict its widespread application [10]. Therefore, there is an urgent need for a more universal and applicable neoadjuvant diagnostic method.

With the development of artificial intelligence (AI), radiomics and deep learning can extract high-dimensional features from medical images, leading to precise predictions. Radiomics can extract high-throughput quantitative image features, while convolutional neural networks (CNNs) in deep learning, based on a hierarchical structure, can recognize subtle differences in images, providing more complex and high-dimensional abstract features [11]. These technologies show great potential in medical image analysis and are expected to overcome the limitations of existing detection methods [[12], [13], [14]]. Compared to mammography and breast ultrasound, breast magnetic resonance imaging (MRI) has shown significant advantages in breast cancer screening. MRI uses magnetic fields and radio waves to generate images such as T1-weighted imaging (T1WI), T2-weighted imaging (T2WI), diffusion-weighted imaging (DWI), and dynamic contrast-enhanced MRI (DCE-MRI), collectively known as multiparametric MRI (mp-MRI), which provides more lesion information and aids clinicians in making more accurate decisions [[15], [16], [17]]. In recent years, the development of automated breast ultrasound (ABUS) technology has expanded the application of ultrasound imaging in breast cancer screening. ABUS automatically acquires ultrasound images and reconstructs them in three dimensions, improving the reproducibility and diagnostic consistency of breast ultrasound [18]. Another study combined ultrasound with mammography by incorporating 3D automated breast ultrasound into mammographic screening for patients with dense breasts, demonstrating a significant improvement in tumor detection rates. This suggests that the integration of multiple imaging modalities can further enhance diagnostic accuracy [19].

Currently, when radiologists perform breast MRI diagnostics, they often need to combine information from multiple imaging modalities, as well as clinical laboratory results and physical examinations. This is because the information provided by different modalities is complementary, and integrating this information contributes to a more comprehensive diagnosis. Previous studies have demonstrated that models integrating multiple sources of information outperform any single modality in diagnostic effectiveness [20]. This finding also highlights the significant advantages of multimodal models in disease diagnosis. In one study [21], a multimodal model that integrates various medical data significantly improved prediction accuracy, reduced physicians' workloads, and increased clinician satisfaction. These results indicate that the integration of multiple modalities can provide richer information, and a more comprehensive analysis of this data can lead to better accuracy. Multimodal imaging not only demonstrates advantages in breast MRI but also plays a crucial role in the diagnosis and treatment evaluation of muscle diseases. Studies have shown that muscle MRI abnormalities correlate with the severity of histopathological changes, with fat infiltration moderately associated with pathological scores (rho = 0.594, p < 0.001) and muscle edema significantly correlated with inflammation severity (p < 0.001). The integration of MRI modalities such as T1, T2, and TIRM allows for a more accurate assessment of muscle disease progression [22]. Additionally, the combined analysis of 34 complementary indicators from mDIXON T1, T1/T2 mapping, and DTI enables quantitative evaluation of thigh muscle abnormalities and aids in preoperative assessment for total knee arthroplasty (TKA) as well as postoperative rehabilitation planning. The fusion of multimodal imaging enhances diagnostic accuracy and provides a basis for personalized treatment [23].

In recent years, multimodal feature fusion networks have been extensively studied and applied in breast cancer research, but the fusion methods vary widely [24]. For example, some studies focus on the simple concatenation of multimodal features, while others apply more complex feature extraction and selection mechanisms. Additionally, factors such as the weight distribution between imaging data and radiomics features, fusion methods, and the architecture design of deep learning networks can also affect the final prediction performance. Therefore, although multimodal feature fusion techniques have shown great promise in classification, optimizing feature fusion methods to improve the model's generalizability across different datasets and clinical scenarios remains a key challenge in this field. This requires not only innovation in methodology but also a thorough exploration of the adaptability of different data sources, features, and algorithms in real-world applications to more accurately distinguish Ki-67 status in breast cancer and provide more reliable decision-support tools for clinical practice. For instance, Zheng et al. [25] constructed a multimodal deep learning system by integrating T1WI, T2WI, DCE-MRI and clinical imaging features to predict breast cancer lymph node metastasis, achieving remarkable results. This study demonstrates the great potential of multimodal information in the clinical application of breast cancer. Similarly, Daimiel Naranjo et al. [26] proposed a multiparametric breast MRI model that combines radiomics and traditional machine learning, which was successfully used for breast lesion classification, with an accuracy of 88.5 % and an AUC of 0.96, further validating the effectiveness of multimodal fusion techniques. However, current research on multimodal approaches for breast cancer Ki-67 status remains limited.

In this context, there is an urgent need to study multimodal models for Ki-67. Combining AI technology with mp-MRI can enable more comprehensive analyses and improve the performance of predicting breast cancer Ki-67 status preoperatively. In recent studies, most models have focused on radiomics [3,27,28], with fewer deep learning models, and even fewer studies have explored multiparametric breast MRI models that integrate radiomics with deep learning. This study aims to establish a breast MRI multimodal model by integrating radiomics and deep learning features from mp-MRI to achieve automatic preoperative prediction of breast cancer Ki-67 status, thereby promoting personalized treatment and precision medicine for breast cancer patients.