Abdominal CT Urography Kidney Segmentation

Retrograde intrarenal surgery is one of a minimally invasive procedures for treatment of renal stones.The anatomy of the kidney and of pelvicalyceal can greatly affect the lithotripsy rate of the procedure.Accurate measurement of these anatomical parameters is important to reduce the risk of retrograde intrarenal surgery.However,accurate parameter measurement requires CTU(Computed Tomography Urography)for 3D segmentation and reconstruction of the kidney.With the increase in the amount of medical image data and the advent of high-performance computing devices,researchers have proposed deep fully convolutional neural networks(FCNNs)and are continuously optimizing them.FCNNs have already achieved satisfactory results for clinicians in tasks such as image segmentation and target detection and recognition.In contrast,traditional algorithms cannot learn rich features from large amount of data efficiently and automatically.Therefore,this paper investigates CTU kidney segmentation based on deep FCNNs to achieve 3D reconstruction of the kidney and thus assists doctors measuring parameters accurately.We firstly found that 3D FCNNs trained from CT data cannot accurately perceive the kidney region in CTU under without CTU kidney annotation.So,we propose a CTU kidney segmentation algorithm based on unsupervised domain translation as well as adversarial learning.The method uses the cycle-consistency loss function and adversarial learning to reduce the difference between CT and CTU data distribution and generate CTU-like data to training a segmentation network.The average Dice coefficient of our method was 0.910 on CTU test data.The experimental results demonstrate that our method can achieve better kidney segmentation results than supervised 3D FCNNs.And our method does not use the kidney annotation information of CTU.Also,its experimental results are higher than the average Dice coefficients of other segmentation methods based on unsupervised domain translation.And then,we found that 2D FCNNs are unable to learn the contextual features between CTU slices effectively,as well as the problem that 2D and 3D FCNNs are prone to overfitting with limited training data.So,we propose a CTU kidney segmentation algorithm based on spatiotemporal full convolutional neural networks.The method uses a deep residual network as the encoder.And its decoder is composed of multiple convolutional LSTM modules with spatiotemporal representation capability.The initialization weights of the encoder are obtained by training natural image data.The average Dice coefficient,average ASD,average false positive rate of our method was 0.946,0.985,and 0.0004838,respectively.The experimental results demonstrate that our method is able to capture the spatial information between CTU slices and effectively mitigate overfitting.