为了更有效地表征肿瘤的生长趋势,提出了基于边缘注入的自适应时空网络(adaptive spatiotemporal net-works based on edge injection,EA-Net).引入紧密型空间增强模块学习多形态肿瘤的静态抽象信息,并重点提取关键生长特征;明确监督目标肿瘤的边缘信息,设计多密度边缘注入模块来增强特征图像内目标的边缘权重;建立广泛的时空信息关联,实现对肿瘤未来成像的自适应建模;引入个体特异性因子,促使网络学习不同人群肿瘤的生长趋势.实验表明,提出的方法能更高效地预测肿瘤未来的生长状态,其Dice score、Recall、RMSE分别可达 89.36%、90.83%、10.42%,能在一定程度上揭示肺癌疾病的发展方向.
Abstract
To more efficiently characterize tumor growth trends,adaptive spatiotemporal networks based on edge injection(EA-Net)are proposed.A compact spatial enhancement module is introduced to learn static abstract information of multi-morphological tumors,focusing on extracting critical growth features.By explicitly supervising the target tumor's edge infor-mation,a multi-density edge injection module is designed to enhance edge weights of the target within feature images.Broad spatiotemporal information associations are established to enable adaptive modeling of future tumor imaging.Individual-spe-cific factors are incorporated to facilitate the network's learning of tumor growth trends across diverse populations.Experi-mental results show that the proposed method can efficiently predict the future growth state of tumors,achieving a Dice score of 89.36%,Recall of 90.83%,and RMSE of 10.42%.This method provides insights into the progression of lung cancer to a certain extent.
万方数据