Deep learning for preoperative MRI-based endometrial cancer staging prediction.

Abstract

Endometrial carcinoma ranks among the most common malignancies of the female reproductive system. Accurate early-stage staging is essential for devising appropriate treatment plans and assessing patient prognosis. This study aims to enhance diagnostic precision by overcoming the limitations of traditional imaging methods and existing deep learning models.To address challenges such as dependency on physician expertise, inefficiency, and deficiencies in feature transmission, boundary detail restoration, and multi-scale feature integration, we propose a novel architecture termed GCMF-UNet (Group Convolution and Multi-Scale Fusion U-Net). Furthermore, for classification tasks, we introduce MSFA-Net (Multi-Scale Fusion Attention Network), which integrates a ResNet-18 backbone with a multi-scale feature aggregation module, squeeze-and-excitation (SE) attention, and a Swin Transformer for global contextual modeling. Experimental results indicate that GCMF-UNet improves Accuracy from 90.1% to 94.2% and Recall from 89.3% to 94.8% compared to the standard U-Net. In classification performance, MSFA-Net improves F1-score from 0.901 to 0.938 over baseline ResNet-18, demonstrating enhanced capability in identifying critical lesion features. The proposed GCMF-UNet and MSFA-Net architectures effectively mitigate limitations of conventional diagnostic and deep learning approaches, offering more accurate lesion segmentation and classification. These advancements offer a technical basis for further exploration of automated diagnosis and staging in endometrial carcinoma.