Liu Wenjuan, Zhang Limin, Li Xiangrui, Liu Haoran, Feng Min, Li Yanxia
Department of Respiratory and Critical Care Medicine, The First Affiliated Hospital of Dalian Medical University, Dalian, 116021, China.
Clinical Medicine, Dalian Medical University, Dalian, 116000, China.
Sci Rep. 2025 Mar 27;15(1):10562. doi: 10.1038/s41598-025-94132-9.
Early screening of lung nodules is mainly done manually by reading the patient's lung CT. This approach is time-consuming laborious and prone to leakage and misdiagnosis. Current methods for lung nodule detection face limitations such as the high cost of obtaining large-scale, high-quality annotated datasets and poor robustness when dealing with data of varying quality. The challenges include accurately detecting small and irregular nodules, as well as ensuring model generalization across different data sources. Therefore, this paper proposes a lung nodule detection model based on semi-supervised learning and knowledge distillation (SSLKD-UNet). In this paper, a feature encoder with a hybrid architecture of CNN and Transformer is designed to fully extract the features of lung nodule images, and at the same time, a distillation training strategy is designed in this paper, which uses the teacher model to instruct the student model to learn the more relevant features to nodule regions in the CT images and, and finally, this paper applies the rough annotation of the lung nodules to the LUNA16 and LC183 dataset with the help of semi-supervised learning idea, and completes the model with the accurate annotation of lung nodules. Combined with the accurate lung nodule annotation to complete the model training process. Further experiments show that the model proposed in this paper can utilize a small amount of inexpensive and easy-to-obtain coarse-grained annotations of pulmonary nodules for training under the guidance of semi-supervised learning and knowledge distillation training strategies, which means inaccurate annotations or incomplete information annotations, e.g., using nodule coordinates instead of pixel-level segmentation masks, and realize the early recognition of lung nodules. The segmentation results further corroborates the model's efficacy, with SSLKD-UNet demonstrating superior delineation of lung nodules, even in cases with complex anatomical structures and varying nodule sizes.
肺结节的早期筛查主要通过人工读取患者的肺部CT来完成。这种方法既耗时又费力，而且容易出现漏诊和误诊。当前的肺结节检测方法面临着诸多限制，比如获取大规模、高质量标注数据集的成本高昂，以及在处理质量各异的数据时鲁棒性较差。挑战包括准确检测小的和不规则的结节，以及确保模型在不同数据源上的泛化能力。因此，本文提出了一种基于半监督学习和知识蒸馏的肺结节检测模型（SSLKD-UNet）。本文设计了一种具有CNN和Transformer混合架构的特征编码器，以充分提取肺结节图像的特征，同时，本文设计了一种蒸馏训练策略，利用教师模型指导学生模型学习CT图像中与结节区域更相关的特征，最后，本文借助半监督学习思想将肺结节的粗略标注应用于LUNA16和LC183数据集，并结合肺结节的准确标注完成模型训练过程。进一步的实验表明，本文提出的模型可以在半监督学习和知识蒸馏训练策略的指导下，利用少量廉价且易于获取的肺结节粗粒度标注进行训练，即不准确的标注或不完整的信息标注，例如使用结节坐标而非像素级分割掩码，实现肺结节的早期识别。分割结果进一步证实了该模型的有效性，SSLKD-UNet在肺结节的描绘方面表现出色，即使在解剖结构复杂且结节大小各异的情况下也是如此。
