上海交通大学学报

上海交通大学学报

• •    下一篇

注意力引导多任务学习的前列腺癌盆腔淋巴结转移预测（网络首发）

  

  1. 1. 徐州医科大学医学影像学院；2. 中国科学院苏州生物医学工程技术研究所；3. 苏州大学附属第二医院
  • 基金资助:
    中国科学院青年创新促进会项目(2021324)；苏州市卫健委临床重点病种诊疗专项（LCZX202001）；苏州市科技计划项目（SKY2022003）

Attention Guided Multi-task Learning for Prostate Cancer Pelvic Lymph Node Metastasis Prediction

  1. 1. School of Medical Imaging, Xuzhou Medical University, Jiangsu, 221000, China；2. Suzhou Institute of Biomedical Engineering and Technology, Chinese Academy of Science, Jiangsu, 215163, China；3. The Second Affiliated Hospital of Suzhou University, Jiangsu, 215000, China

PDF (PC)

99

摘要： 基于前列腺癌原发灶的术前磁共振影像定量特征预测盆腔淋巴结转移（pelvic lymph node metastasis， PLNM）是治疗方案制定的重要参考。然而，现有预测方法对肿瘤原发灶内部的异质性信息提取不充分，导致提取的图像定量特征与PLNM预测的关联性不足。针对以上问题，提出一种肿瘤分割任务作为辅助任务的注意力引导多任务学习网络用于PLNM预测。首先，在肿瘤分割网络中提出多分支各向异性大核注意力模块，通过不同分支、各向异性大卷积核获得更大的感受野从而有效获取肿瘤的局部和全局信息。然后，在PLNM预测网络中提出多尺度特征交互融合注意力模块，对多尺度特征进行层次化融合筛选。在320例数据集上的实验显示，所提方法的AUPRC值（85.44%±2.04%）和AUROC值（91.86%±2.18%）优于经典的单任务分类方法和多任务方法。

关键词: 前列腺癌盆腔淋巴结转移, 多任务学习, 多分支各向异性大核注意力模块, 多尺度特征交互融合注意力模块, 多参数磁共振

Abstract: Prediction of pelvic lymph node metastasis (PLNM) based on quantitative preoperative magnetic resonance imaging features of prostate cancer primary tumor is an important reference for treatment planning. However, current prediction methods inadequately capture the heterogeneity information within the primary tumor, resulting in insufficient correlation between extracted quantitative image features and PLNM prediction. In response to the aforementioned issues, an attention guided multi-task learning network with tumor segmentation as an auxiliary task is proposed for PLNM prediction. Firstly, within the tumor segmentation network, a multi-branch anisotropic large kernel attention module is introduced, where a larger receptive field is obtained through different branches and anisotropic large convolutional kernels, effectively capturing both local and global tumor information. Subsequently, within the PLNM prediction network, a multi-scale feature interaction fusion attention module is introduced to hierarchically fuse and select features from multiple scales. Experimental results on a dataset of 320 cases demonstrate that the proposed method achieves superior AUPRC (85.44%±2.04%) and AUROC (91.86%±2.18%) values compared to state-of-the-art methods and multi-task methods.

Key words:

中图分类号: