Research on the diagnostic value of ultrasound combined with MRI features based on deep learning in distinguishing benign and malignant ovarian tumors

1. Introduction

Ovarian tumors are one of the most common tumors of the female reproductive system, with malignant types (ovarian cancer) often referred to as the “king of gynecological cancers” due to their insidious onset and high aggressiveness [1]. Globally, more than 300,000 new cases of ovarian cancer are diagnosed each year, and its mortality rate ranks first among gynecological malignancies. The 5-year survival rate of Chinese patients remains below 50%, which is significantly lower than that of developed countries [2]. Clinical data show that approximately 70% of ovarian cancer patients are diagnosed at an advanced stage (Ⅲ–Ⅳ), when tumor metastasis to the pelvic and abdominal cavity has already occurred, resulting in a sharp decline in the 5-year survival rate to below 30%. In contrast, early-stage (I) patients can achieve a 5-year survival rate exceeding 90% after standardized treatment [3, 4]. Therefore, early and accurate differentiation between benign and malignant ovarian tumors is crucial for improving patient prognosis.

Currently, the preoperative diagnosis of ovarian tumors primarily relies on imaging examinations. Ultrasound (US), being radiation-free, real-time, convenient, and cost-effective, serves as a first-line screening modality [5]. By evaluating tumor characteristics, such as size, morphology, wall structure, septa, internal echogenicity, and vascularity, experienced physicians can make a preliminary assessment for most tumors [6, 7]. Magnetic Resonance Imaging (MRI), known for its superior soft tissue resolution and multi-sequence imaging capabilities, plays a vital complementary role in the differential diagnosis of indeterminate lesions, identification of tumor origin, and preoperative staging [8, 9]. However, traditional imaging diagnosis remains highly dependent on physicians’ experience and subjective visual assessment, leading to intra- and inter-observer variability. For atypical or complex cases, even the combined use of US and MRI may not achieve optimal diagnostic accuracy [10]. With the rapid development of artificial intelligence (AI), deep learning (DL), its core branch, has made revolutionary contributions to medical image analysis [11]. Convolutional Neural Networks (CNNs) can automatically learn and extract multi-level, high-dimensional features from large-scale imaging data that are often imperceptible to the human eye, enabling accurate disease detection, segmentation, and classification [12, 13]. Currently, the application of DL algorithms in the diagnosis and prognosis prediction of gynecological tumors is mainly focused on cervical cancer, followed by ovarian and endometrial cancers [14]. Meta-analyses have confirmed that AI- and DL-based imaging features possess high diagnostic value for ovarian tumors [15]. Previous studies have demonstrated that DL models analyzing either US or MRI images alone outperform traditional manual assessments, establishing a new paradigm for computer-aided diagnosis [16, 17, 18, 19]. Integrating DL-derived features from both US and MRI modalities is expected to yield a more comprehensive and robust diagnostic framework, thereby achieving a “1 + 1 > 2” synergistic effect in diagnostic performance.

Therefore, this study aimed to investigate the diagnostic efficacy of a DL-based multimodal model combining US and MRI features for differentiating benign from malignant ovarian tumors, and to develop an intelligent auxiliary diagnostic tool for clinical practice.

2. Materials and methods

2.1 Study subjects

This study was a retrospective cohort analysis. Data were collected from patients with ovarian tumors treated at Shanxi Bethune Hospital from January 2022 to December 2024. Based on preliminary pilot study results (AUC of ultrasound alone: 0.82; expected AUC of the combined model: 0.95; α = 0.05; β = 0.2), the required sample size was calculated using MedCalc 19.0 software (MedCalc Software Ltd, Ostend, Belgium). A total of 900 patients were initially considered; after applying the inclusion and exclusion criteria, 887 patients were finally enrolled, including 665 with benign tumors and 222 with malignant tumors. In the binary classification task, borderline tumors were classified as malignant tumors. The detailed patient selection flowchart is shown in Fig. 1. This study was approved by the Institutional Ethics Committee of Shanxi Bethune Hospital (approval number: 2025-356). All patients provided written informed consent, and the study complied with the Declaration of Helsinki.

Fig. 1.Flowchart of the study population selection process.

2.4 Deep learning model construction

A dual-channel multimodal fusion model was adopted (Fig. 2), comprising three main components:

Fig. 2.Schematic diagram of the deep learning multimodal fusion model architecture. US: ultrasound; MRI: magnetic resonance imaging; ZERO PAD: Zero Padding; CONV: Convolutional Layer; Batch Norm: Batch Normalization; ReLU: Rectified Linear Unit; MAX POOL: Max Pooling Layer; ID: Identity; AVG POOL: Global Average Pooling; FC: Fully Connected Layer; ResNet50: Residual Network with 50 layers. *: Sigmoid activation function.

1. Unimodal Feature Extraction Modules: The unimodal feature extraction module consists of an ultrasound branch and an MRI branch. The ultrasound branch is based on the ResNet50 architecture, with inputs of 2D ultrasound images (grayscale + CDFI fused images), and extracts morphological features (edges, textures) and blood flow features (signal intensity distribution) through convolutional layers. The MRI branch is based on the 3D ResNeXt101 architecture, with inputs of MRI multi-sequence fused data (T2WI + DCE-MRI + ADC maps), and extracts spatial features (intratumoral structural heterogeneity) and functional features (degree of diffusion restriction, blood perfusion) through 3D convolutions.

2. Cross-modal Fusion Module: The cross-modal fusion module employs an attention mechanism fusion (Attention Fusion). It performs dimension alignment between the 2D feature maps (512 channels) output by the US branch and the 3D feature maps (1024 channels) output by the MRI branch (reducing 3D features to 2D via global average pooling); calculates modal attention weights (Eqn. 1), followed by weighted feature fusion (Eqn. 2):

$W_u=Softmax(FC(GAP(F_u))),W_m=Softmax(FC(GAP(F_m)))$(1)

$F_{fusion}=W_u\cdot F_u+W_m\cdot F_m$(2)

In Eqn. 1: W_u: ultrasound modal attention weight; F_u: ultrasound features; W_m: MRI modal attention weight; F_m: MRI features; GAP: global average pooling; FC: fully connected layer.

3. Classification module: After the fused features pass through two fully connected layers (with Dropout = 0.5), the predicted probabilities of benign and malignant tumors (benign = 1, malignant = 2) are output through the Softmax function.

3. Results

3.3 ROC curve analysis

The ROC curve analysis demonstrated that the DL multimodal fusion model achieved the highest area under the curve (AUC = 0.957) for differentiating benign and malignant ovarian tumors. The AUC was significantly higher than that of US alone (z = 3.92, p < 0.001), MRI alone (z = 2.76, p = 0.006), and the ADNEX model (z = 2.07, p = 0.022). Details are shown in Table 5 and Fig. 3.

Table 5.Comparison of AUC values of different diagnostic methods.

Diagnostic method	AUC	Standard error	p value	95% confidence interval
Ultrasound diagnosis alone	0.792	0.041	<0.001	0.712–0.872
MRI standalone diagnosis	0.844	0.035	<0.001	0.774–0.913
ADNEX model diagnosis	0.885	0.032	<0.001	0.823–0.948
DL model diagnosis	0.957	0.021	<0.001	0.915–0.998
MRI: magnetic resonance imaging; ADNEX: Assessment of Different NEoplasias in the adnexa; DL: deep learning; AUC: Area Under the Curve.

Fig. 3.ROC Curves of Different Diagnostic Methods. MRI: magnetic resonance imaging; ADNEX: Assessment of Different NEoplasias in the adnexa; DL: deep learning.

4. Discussion

As the “king of gynecological cancers”, early diagnosis of ovarian cancer is crucial for improving patient prognosis. This study developed a deep learning (DL)-based multimodal fusion model that combines ultrasound (US) and MRI to enhance the accuracy of differentiating between benign and malignant ovarian tumors. By automatically extracting and fusing high-dimensional imaging features from both modalities, the model achieved significantly higher diagnostic performance than US alone, MRI alone, or the ADNEX model, thereby demonstrating the substantial advantage of multimodal deep learning in ovarian tumor diagnosis.

The findings of this study indicate that combining the DL-extracted US and MRI features can significantly improve diagnostic accuracy for distinguishing benign from malignant ovarian tumors. Using Convolutional Neural Networks (CNN), the model automatically identifies latent high-dimensional features in US and MRI images, effectively compensating for the limitations of traditional visual assessment [20, 21]. Compared with the limitation of conventional imaging diagnosis relying on physicians’ subjective experience, this model effectively reduces inter-observer variability by automatically extracting deep features that are difficult for the human eye to recognize, such as ultrasound probability maps of papillary projections and MRI DCE-MRI time-signal curve types. It particularly exhibits better recognition ability for atypical cases, such as mixed tumors and small papillary projections [22, 23].

From a clinical practice perspective, the high sensitivity of the model can reduce the missed diagnosis rate of malignant tumors, while the high specificity can reduce over-treatment of benign tumors, which is of great significance for avoiding unnecessary surgical trauma and optimizing treatment decisions [24]. For patients with International Federation of Gynecology and Obstetrics (FIGO) stage Ⅰ, early diagnosis can increase the 5-year survival rate to more than 90%, and the high diagnostic performance of this model is expected to provide support for timely intervention in more early-stage cases [25].

This study introduced an innovative dual-channel fusion architecture. The 2D residual architecture of ResNet50 was optimized for US 2D planar data, efficiently extracting morphological and hemodynamic features through residual connections and moderate depth [26]. The 3D ResNeXt101 branch, designed for volumetric MRI data, captured spatial–functional information using grouped convolutions and high cardinality. An attention-based fusion mechanism dynamically weighted the contributions of both modalities, ensuring optimal integration. The superior AUC of the combined model confirmed the complementarity of multimodal information [27]. By integrating the real-time vascular assessment capability of ultrasound with superior soft-tissue contrast and functional imaging power of MRI (DWI, DCE), this model provides a comprehensive reflection of the tumor biology, including blood supply, cell density, and perfusion patterns [28]. This fusion strategy offers a novel approach to addressing the problem of limited information from a single modality, and can serve as a framework for other multimodal medical image applications, such as breast and liver tumor diagnosis [29]. From a clinical translation perspective, the model could be integrated into hospital Picture Archiving and Communication Systems (PACS) workflows to provide real-time auxiliary diagnostic assistance during radiologists’ image review. This would reduce the inter-observer variability and shorten reporting time. For cases with high Ovarian-Adnexal Reporting and Data System (O-RADS) scores (4–5) identified during the initial US screening, the system could combine the MRI-derived features to provide rapid malignancy risk stratification, supporting timely surgical or follow-up decisions and streamlining diagnostic workflows.

Recent studies have investigated the application of DL in ovarian tumor imaging, yet most have focused on single-modality approaches. For example, US-based CCN models have reported AUCs around 0.941 [30], while an MRI-based DL model achieved an AUC of 0.89 [31], both lower than the combined model in this study (0.957). The AUC of conventional combined diagnosis typically ranges from 0.85 to 0.90 and is limited by inter-physician experience variability [32, 33, 34]. In contrast, our automated feature fusion approach eliminated subjective bias and revealed latent cross-modality relationships, achieving a notable improvement in diagnostic accuracy.

Compared with the ADNEX model, our DL model does not rely on serum markers, yet achieves higher accuracy using only imaging data, making it particularly suitable for early-stage patients with normal CA125 levels or atypical cases [35, 36]. This advantage expands the model’s application scenarios and reduces reliance on laboratory tests. Furthermore, Adusumilli et al. [37] recently proposed a PyTorch- Medical Open Network for AI framework incorporating multiple architectures (ResNet, Densely Connected Convolutional Network, Transformer-based Unified Efficient Segmentation Transformer, and attention-based Multiple Instance Learning). Although their design is more comprehensive, it remains conceptual, whereas our implemented ResNet-based model provides direct experimental validation and demonstrated clinical feasibility.

Despite the significant findings of this study, the following limitations should be acknowledged: (a) This was a single-center retrospective cohort study, and sample selection may be subject to bias. The generalizability of the model requires validation with multi-center data. (b) The US and MRI data were acquired from specific models of equipment. Variations in image quality and parameter settings across different brands of devices may affect the model’s generalizability, and cross-device compatibility issues need to be addressed for practical application. (c) The study included benign, borderline, and malignant tumors; however, the diagnostic performance for borderline tumors was not analyzed separately. Borderline tumors exhibit biological behavior intermediate between benign and malignant, with more complex imaging features. Future studies should specifically optimize the model’s recognition ability for this subgroup.

To address the above limitations, future research can be advanced in the following directions: (a) Conduct multi-center, prospective studies incorporating data from different devices and diverse populations (e.g., young women, pregnant patients) to validate the model’s stability and universality and explore its feasibility for deployment in primary hospitals. (b) Incorporating additional data modalities (e.g., Computed Tomography, Positron Emission Tomography-Computed Tomography) and clinical biomarkers (e.g., Human Epididymis Protein 4, miRNA profiles) to enhance diagnostic accuracy through multi-omics fusion. (c) Implementing advanced data harmonization techniques and exploring federated learning or domain adaptation to improve model generalization in real-world clinical environments. (d) Developing lightweight, mobile-compatible DL models that can integrate with PACS workflows, enabling real-time decision support for radiologists during patient evaluation.

5. Conclusions

This study demonstrates that a deep learning-based multimodal fusion model integrating US and MRI can significantly improve the diagnostic performance in differentiating benign and malignant ovarian tumors. The proposed deep learning model shows strong potential for clinical translation as an intelligent “second opinion” system to assist radiologists. By integrating it into hospital PACS workflows, this model could enhance diagnostic efficiency, reduce reporting time, and guide MRI utilization for suspicious cases, ultimately optimizing patient management and improving clinical outcomes.

Availability of data and materials

The authors declare that all data supporting the findings of this study are available within the paper and any raw data can be obtained from the corresponding author upon request.

Author contributions

NS—designed the study and carried them out. NS, XLY, LLF—supervised the data collection; analyzed the data; interpreted the data. NS, NZ, YX—prepared the manuscript for publication and reviewed the draft of the manuscript. All authors have read and approved the manuscript.

Ethics approval and consent to participate

Ethical approval was obtained from the Ethics Committee of Shanxi Bethune Hospital (Approval No. 2025-356). Written informed consent was obtained from legally authorized representatives for anonymized patient information to be published in this article.

Acknowledgment

Not applicable.

Funding

This research received no external funding.

Conflict of interest

The authors declare no conflict of interest.