上海科技大学知识管理系统

Accurate detection and quantification of lymph nodes from contrast-enhanced CT images are crucial for cancer diagnosis, staging, treatment planning, and prognosis. Detecting lymph nodes in the mediastinal area
is challenging due to low contrast, irregular shapes, and dispersed distribution of the nodes. This paper proposes a hybrid detection network, SDF-Net+, which integrates segmentation and detection models to enhance lymph node detection. To achieve a deeper integration of the two networks, we introduce a Feature Fusion (FF) module. This module effectively merges feature maps from a segmentation model (Seg-path) into a detection model (Det-path). By incorporating detailed lymph node morphology, the FF module enhances detection accuracy and accelerates model convergence. Then, SDF-Net+ introduces a NonLinear Fusion (NLF) module to nonlinearly fuse the probability maps (PBM) and feature maps from both paths at the final level. During the model training, a three-dimensional (3D) shape-adaptive Gaussian kernel is introduced to represent a lymph node. This kernel provides more anatomical information and effectively facilitates the learning process without requiring precise mask annotations. Experiments conducted on a dataset of 1070 contrast-enhanced CT images with lymph node annotated by 3D bounding boxes. The results show that the proposed SDF-Net+ achieves overall best performance in recall, accuracy, and F1-score, compared with the state-of-the-art methods. The code is available at https://github.com/xiongjiuli/Lymph-Nodes-3DDetction.