This study aimed to propose an effective malignant solitary pulmonary nodule classification method based on improved Faster R-CNN and transfer learning strategy. In practice, the existing solitary pulmonary nodule classification methods divide the lung cancer images into two categories only: normal and cancerous. This study proposed the deep convolution neural network to classify the computed tomography (CT) images of lung cancer into four categories: lung adenocarcinoma, lung squamous cell carcinoma, metastatic lung cancer, and normal types of lung cancer. Some high-resolution lung CT images have unnecessary characters such as a large number of high-density continuity features, small-size lung nodule targets, CT image background complexity, and so forth. In this study, the CT image sub-block preprocessing strategy was used to extract nodule features for enhancement and alleviate the aforementioned problems. The experimental results showed that the proposed system was effective in resolving issues such as high false-positive rate and long classification time cost based on the original Faster R-CNN detection method. Meanwhile, the transfer learning strategy was used to improve the classification efficiency so as to avoid the overfitting problem caused by a few labeled samples of lung cancer datasets. The classification results were integrated using the majority vote algorithm. The classification results of the lung CT imaging showed that the proposed method had an average detection accuracy of 89.7% and reduced the rate of misdiagnosis to meet the clinical needs.

中文翻译：

---

基于转移学习策略的卷积神经网络对孤立性肺结节的自动分类

本研究旨在提出一种基于改进的 Faster R-CNN 和迁移学习策略的有效恶性孤立性肺结节分类方法。在实践中，现有的孤立性肺结节分类方法仅将肺癌图像分为两类：正常和癌性。本研究提出深度卷积神经网络将肺癌的计算机断层扫描 (CT) 图像分为四类：肺腺癌、肺鳞状细胞癌、转移性肺癌和正常类型的肺癌。一些高分辨率肺部CT图像具有大量高密度连续特征、小尺寸肺结节目标、CT图像背景复杂等不必要的特征。在这项研究中，CT图像子块预处理策略用于提取结节特征进行增强并缓解上述问题。实验结果表明，该系统在原有Faster R-CNN检测方法的基础上，有效解决了误报率高、分类时间成本长等问题。同时，采用迁移学习策略提高分类效率，避免肺癌数据集标记样本过少造成的过拟合问题。使用多数投票算法对分类结果进行整合。肺部CT影像分类结果表明，该方法平均检测准确率为89.7%，降低误诊率，满足临床需要。