LUNA16

Lung cancer is the most prevalent and deadly oncological disease in the world, but a timely detection of lung nodules can greatly improve the survival rate of this disease. However, due to the tiny size of lung nodules and inconspicuous edges, lung nodules are not easily distinguished by naked eyes thus medical image diagnosticians are prone to misdiagnosis simply based on their own experiences and subjective judgements. In recent years, the machine-learning-based image processing techniques find their wide applications in the field of medical diagnosis, and have been proved to be an efficient way to aid diagnosticians to accurately identify subtle lesions in images. To accurately recognize lung nodules in CT images, in this paper, we propose an approach, called STBi-YOLO. This approach stems from YOLO-v5, but makes significant improvements from three dimensions—we first use the spatial pyramid pooling network that is based on stochastic-pooling method to modify the basic network structure of YOLO-v5; then apply a bidirectional feature pyramid network to perform multi-scale feature fusion; finally improve the loss function of the YOLO-v5 and adopt the EIoU function to optimize the training model. To evaluate our approach, we compare STBi-YOLO with YOLO-v3, YOLO-v4, YOLO-v5, and multiple leading object detection models, such as Faster R-CNN and SSD. The experiments show that STBi-YOLO achieves an accuracy of 96.1% and a recall rate of 93.3% for the detection of lung nodules, while producing a 4× smaller model size in memory consumption than YOLO-v5 and exhibiting comparable results in terms of mAP and time cost against Faster R-CNN and SSD.

肺癌是全球范围内发病率最高、致死性最强的肿瘤疾病，及时检出肺结节可显著提升该疾病的患者生存率。然而，由于肺结节体积微小、边缘特征不明显，肉眼难以对其进行分辨，仅依靠医学影像诊断医师的个人经验与主观判断极易出现误诊。近年来，基于机器学习的图像处理技术在医学诊断领域得到广泛应用，被证实为辅助医师精准识别影像中细微病变的高效途径。为实现CT影像中肺结节的精准识别，本文提出一种名为STBi-YOLO的检测方法。该方法以YOLO-v5为基础框架，从三个维度进行了重大改进：首先采用基于随机池化（stochastic-pooling）的空间金字塔池化网络（spatial pyramid pooling network）对YOLO-v5的基础网络结构进行改良；其次应用双向特征金字塔网络（bidirectional feature pyramid network）完成多尺度特征融合；最后对YOLO-v5的损失函数进行优化，采用EIoU函数对训练模型进行调优。为验证所提方法的有效性，本文将STBi-YOLO与YOLO-v3、YOLO-v4、YOLO-v5以及Faster R-CNN、SSD等多款主流目标检测模型开展对比实验。实验结果表明，STBi-YOLO在肺结节检测任务中准确率达96.1%、召回率达93.3%，模型内存占用量仅为YOLO-v5的1/4，且在平均精度均值（mean Average Precision, mAP）与时间开销方面与Faster R-CNN、SSD表现相当。