结合Sobel算子和Mask R-CNN的肺结节分割

肺癌不断威胁着人类健康,计算机辅助诊断对肺癌诊断将发挥重要的作用.卷积神经网络(CNNs)在对图像的处理上表现出有目共睹的优秀性能,医学Computed Tomography(CT)图像是用来诊断肺癌的主要检查方式,用深度学习分割病灶的方法可以实现端对端的辅助诊断,这将节省医生的诊断时间,为患者争取最佳治疗时间.LIDC-IDRI(The Lung Image Database Consortium)数据集影像中的癌症部分与其他组织部分的放射密度十分接近,而且往往癌症部分非常小,背景具有非常强的相似性.本文使用传统的Sobel算子对图像中放射密度高的部分进行边缘锐化处理,用强化边缘特征的方法解决前景与背景灰度相似的问题,然后在使用传统的分割方法--阈值分割进一步强化.本文减小Regions of Interest(RoIs)的大小以适应肺结节的特征,减少RoIs的个数以避免过多的负类样例训练产生退化的模型;在传统图像增强处理方法和深度学习的结合下,获得了一个优化的Mask R-CNN模型,在LIDC-IDRI数据集上的测试结果中,基于Intersection over Union(IoU)=0.5的标准下的肺结节平均精度mAP达到72.2%,在FPR为0.226时的TPR达到0.915.

Lung Nodule Segmentation Combining Sobel Operator and Mask R-CNN

The lung cancer is a serious threat to the health of human,computer-aided diagnosis(CAD)will play an increasingly important role in the diagnosis of lung cancer.Convolutional neural networks(CNNs)showoutstanding performance in the processing of images.Computed Tomography(CT)images are the main methods for diagnosis of lung cancer.The method of deep learning to segment lesions can achieve end-to-end auxiliary diagnosis,which will save the doctor’s diagnosis time and get the best treatment time for the patient.The cancer part in the LIDC-IDRI(The Lung Image Database Consortium)dataset has a very close radiation density to the other tissue parts,and usually the cancer part is very small and has a very strong similarity to its surroundings.In this paper,the traditional Sobel operator is used to sharpen the contour of high-density part to solve the problem that the foreground gray is similar to the background,and then further strengthened by the traditional threshold-based segmentation method.In this paper,the size of Regions of Interest(RoIs)are reduced to adapt to the characteristics of pulmonary nodules,the number of RoIs are reduced to avoid excessive negative examples’training;after the combination of traditional image processing methods and deep learning,an optimized Mask R-CNN model is obtained.We test the results on the LIDC-IDRI dataset.The mean average accuracy of lung nodules based on Intersection over Union(IoU)=0.5 is 72.2%,and when FPR=0.226,TPR achieves 0.915.