基于多尺度和特征融合的肺癌识别方法

针对病人肺结节大小各异、结节征象复杂造成的结节检测困难问题,基于迁移学习提出一种多尺度和特征融合的肺癌识别方法,根据CT图像预测病人未来一年内患肺癌的概率。根据肺结节和肺肿块大小,采用3种不同尺度的图像块输入三维结节检测网络,避免小尺度输入的结节检测网络难以获取大区域病灶整体特征的问题;在多尺度输入基础上采用特征融合策略,将网络提取的瓶颈层特征和输出层特征融合,充分描述病灶的详细特征。在Kaggle Data Science Bowl 2017数据集上的实验结果表明,所提方法降低了肺癌预测的损失值,提高了肺癌识别精度。

Lung cancer recognition method based on multi-scale and feature fusion

Aiming at the difficulty in detecting nodules caused by different sizes of lung nodules and complex nodule signs,a lung cancer recognition method based on transfer learning was proposed,in which multi-scale and feature fusion were combined.The method predicted the probability of a patient having lung cancer for the next year according to CT images.Three different scale image blocks were used to input the three-dimensional nodule detection network according to the size of lung nodules and lung masses,which avoided the problem that it is difficult for the small-scale input nodule detection network to obtain the overall features of large-area lesions.The feature fusion strategy was adopted to fuse the bottleneck layer features and output layer features of the network to fully describe the detailed features of the lesion on the basis of multi-scale input.Experimental results on the Kaggle Data Science Bowl 2017 dataset show that the proposed method reduces the predicted loss of lung cancer and improves the accuracy of lung cancer recognition.