融合形状变换及纹理学习的肺结节生长预测

Predicting Lung Nodule Growth with Shape Transformation and Texture Learning

虽然人工智能在肺结节检测方面已经相当成熟,但对其生长预测的研究仍然有限。准确的生长预测有助于临床决策,为患者随访策略提供信息。本文提出一种新的结节生长预测网络模型,该模型可以在特定时间间隔生成高质量的肺结节图像。模型使用双分支结构对肺结节图像进行特征提取,其中一个分支,利用位移场预测机制,通过体素水平的未来位移估计来学习肺结节的形状转换;另一分支,采用3D U-Net,学习肺结节的纹理变化。随后,对提取的高维特征图通过坐标注意力机制,突出有利的图像特征,再拼接两个分支的结果,输入至特征重建模块得到最终的肺结节生长预测图像。同时,本文引入时间间隔编码模块,将期望的时间间隔纳入网络,从而能够生成不同未来时间点的预测图像。

While artificial intelligence has achieved considerable maturity in lung nodule detection,research on growth prediction remains limited.Accurate growth prediction aids clinical decision-making,informing patient follow-up strategies.This paper proposes a novel nodule growth prediction network model that generates high-quality lung nodule images at specific time intervals.The model employs a two-branch structure for feature extraction.One branch,leveraging a displacement field prediction mechanism,models the shape transformation of pulmonary nodules through voxel-level future displacement estimation.The other branch,empowered by a three-dimensional U-Net,focused on learning texture changes within the nodules.A coordinate attention mechanism that emphasizes informative features within the extracted high-dimensional feature map.Subsequently,the outputs of both branches are fused and fed into the feature reconstruction module to generate the final lung nodule growth prediction image.Furthermore,a time interval coding module is introduced to incorporate the desired time interval into the network,enabling the generation of prediction images for different future time points.