Deep Learning-based Magnetic Resonance Imaging Artificial Intelligence Classification Technology in the Diagnosis of Endometrial Cancer

Precise preoperative risk stratification for endometrial cancer is crucial for individualized treatment. Traditional magnetic resonance imaging (MRI) diagnosis is highly subjective, and existing deep learning models have limitations in fusing multi-sequence information and co-modeling features. This study proposes a new Multi-modal Deep Fusion Network (MDFNet) to achieve automatic low-risk/high-risk classification of endometrial cancer based on multi-sequence MRI. This study retrospectively enrolled preoperative multiparametric MRI data from 220 patients with endometrioid adenocarcinoma, wherein 168 cases constituted the training/validation set and 52 cases formed the independent test set, with the dataset strictly divided in chronological order to ensure the authenticity and reliability of the test results. Risk stratification was performed strictly in accordance with the 2025 ESGO-ESTRO-ESP (European Society of Gynaecological Oncology - European Society for Radiotherapy & Oncology - European Society for Pathology) guidelines, and the dataset was divided in chronological order. Experimental results showed that MDFNet achieved excellent classification performance on the independent test set, with an area under the curve (AUC) of 0.912. The accuracy, precision, recall, and F1-score were 0.865, 0.857, 0.864, and 0.860, respectively, which were significantly superior to those of traditional radiomics models and single-path deep learning models across all key evaluation metrics. Ablation experiments verified the effectiveness of the dual-path structure and the attention fusion mechanism. Attention weight analysis revealed that the model's decision-making pattern, which focused more on local features for high-risk tumors, was consistent with clinicopathological logic. Gradient-weighted class activation mapping visualization further demonstrated that the regions of interest highly overlapped with key anatomical landmarks, indicating favorable interpretability. The MDFNet model proposed in this study can effectively integrate multi-sequence MRI information to achieve accurate and interpretable preoperative risk stratification, providing a valuable intelligent auxiliary tool for clinical decision-making. Future work needs to further validate its generalization ability and clinical utility through multi-center, prospective studies.