Weakly-supervised Convolutional Neural Networks For Kidney Segmentation In Abdominal CTA Images

Kidney cancer is one of the ten most common cancers in human beings,and surgery is the best choice for the treatment of kidney cancer and other benign tumors of the kidney.Among them,nephron sparing surgery(NSS)is an effective method for the treatment of kidney tumors.Before NSS surgery,Computed tomography angiography(CTA)needs to determine the information such as location and size.Therefore,it’s necessary to locate the kidney in the CTA image and draw the contour of the relevant tissues,which is an important step in the NSS surgical plan.By constructing the 3d model of the kidney,it can provide a lot of important information for doctors to make an accurate preoperative surgical plan.In recent years,with the development of deep learning technology,convolution neural network(CNN)can be used to generate precise pixel-wise segmentation of the renal tumors in CTA images.However,constructing the training dataset with a large amount of pixel-wise annotations is a time-consuming task for the radiologists.Therefore,weakly-supervised methods have attracted more interest in research,which do not need pixel-wise labels and aim to use simple masks achieve the similar result to fully-supervised methods.We proposed a weakly-supervised method based on the iterative CNN training to obtain accurate segmentation from the bounding box of kidney tumor.First,the Conditional Random Field(CRF)was used to generate the initial label from bounding box.Then,the initial label was used for iterative CNN training.In each training,the output of CNN was used as the unary potential of the CRF model,and the adjusted result was used as the label of the next training.When the training result appears two consecutive drops or reaches the set number to terminate the training.Clinical abdominal CT angiographic images of200 patients were applied to perform the evaluation.Extensive experimental results show that the performance of the iterative training method in the dice coefficient(0.765)was better than that of the traditional one-time training method(0.737).However,the method still has some problems such as imprecise segmentation and long time of training.To improve the above shortcomings,we propose a new weakly-supervised method based on grouping-fusion CNN training.First,the Convolutional Conditional Random Field(Conv CRF)was used to generate the initial weak label from the bounding box.Then,the training dataset was divided into several sub-training datasets which was trained separately by the CNN model.The multiple models were used to generate multiple maps for training dataset,and the weighted sum of these maps was used to generate fusion labels.Finally,the fusion labels and constrained loss function were used for fusion training.Grouping-fusion training can reduce the over-fitting of the improper segmentation in the weak labels during network training and learn the common features of the tumor from multiple models.Therefore,it can produce more reliable segmentation results.In addition,compared with iterative training,the training time of grouping-fusion training is greatly reduced.Experimental results show that the proposed approach achieves the high dice coefficient of 0.834,which is close to fully-supervised method(0.859).