基于多模态医学影像的卵巢癌早期智能筛查模型-《中国医学物理学杂志》

文章信息/Info

Title:: Intelligent early screening model of ovarian cancer based on multimodal medical images

作者:: 高鹏帅1; 2; 刘梦迪2; 樊冬梅1; 1.河南科技大学第一附属医院妇科肿瘤病区，河南洛阳 471003；2.河南科技大学临床医学院，河南洛阳471023

Author(s):: GAO Pengshuai1; 2; LIU Mengdi2; FAN Dongmei1; 1. Gynecological Oncology Ward, the First Affiliated Hospital of Henan University of Science and Technology, Luoyang 471003, China 2. School of Clinical Medicine, Henan University of Science and Technology, Luoyang 471023, China

Keywords:: Keywords: multimodal medical image ovarian cancer early screening deep learning convolutional neural network attention mechanism image fusion

摘要:: 为解决卵巢癌早期病灶隐匿、临床筛查准确率低的问题，提升早期筛查精度以改善患者预后，构建高效、精准的卵巢癌早期智能筛查模型。研究采用多模态医学影像作为核心输入，优先融合磁共振成像（MRI）与超声两种互补性模态，对影像进行格式统一、去噪增强、病灶分割、特征标准化等预处理后，结合深度学习算法搭建智能筛查模型。模型通过卷积神经网络（CNN）分别提取超声与MRI影像的深层特征，采用注意力机制（Attention Mechanism）实现特征级融合，借助Softmax分类器完成“正常卵巢-良性病变-早期卵巢癌”3分类预测。模型结果显示，其对早期卵巢癌的筛查灵敏度、特异性、准确率及ROC曲线下面积（AUC值）均显著优于单一模态筛查模型，可有效减少漏诊、误诊情况。该模型可为卵巢癌早期临床筛查提供可靠的智能化技术支撑，具备良好的临床应用价值与推广前景。

Abstract:: A high-efficiency and accurate intelligent early screening model for ovarian cancer is established to address the problems of hidden early lesions and low clinical screening accuracy, thereby improving early screening precision and patient prognosis. This study used multimodal medical images as core inputs, and prioritizes the fusion of two complementary modalities: magnetic resonance imaging (MRI) and ultrasound. Following image preprocessing, including format unification, noise removal and enhancement, lesion segmentation, and feature standardization, a deep learning-based intelligent screening model is constructed. This model extracts deep features of ultrasound and MRI images separately by convolutional neural networks, realizes feature-level fusion using an attention mechanism, and completes a three-category prediction of normal ovary, benign lesion and early ovarian cancer through a Softmax classifier. Results demonstrated that the proposed model outperforms single-modal screening models in terms of sensitivity, specificity, accuracy and area under the ROC curve for early ovarian cancer screening, effectively reducing missed diagnosis and misdiagnosis. The proposed model can provide reliable intelligent technical support for early clinical screening of ovarian cancer, demonstrating favorable clinical application value and promotion prospects.