Kidney and Kidney Tumor Segmentation

Kidney cancer is one of the most common malignancies worldwide. Early diagnosis is an effective way to reduce the mortality and automated segmentation of kidney tumor in computed tomography scans is an important way to assisted kidney cancer diagnosis. In this paper, we propose a convolution-and-transformer network (COTRNet) for end to end kidney, kidney tumor, and kidney cyst segmentation. COTRNet is an encoder-decoder architecture where the encoder and the decoder are connected by skip connections. The encoder consists of four convolution-transformer layers to learn multi-scale features which have local and global receptive fields crucial for accurate segmentation. In addition, we leverage pretrained weights and deep supervision to further improve segmentation performance. Experimental results on the 2021 kidney and kidney tumor segmentation (kits21) challenge demonstrated that our method achieved average dice of 61.6%, surface dice of 49.1%, and tumor dice of 50.52%, respectively, which ranked the \(22_{th}\) place on the kits21 challenge.

In this paper we present our approach for the KiTS21 Challenge. The goal is to automatically segment kidneys, (renal) tumors and (renal) cysts based on 3D computed tomography (CT) images of the abdomen. The challenge provided public training 300 cases for this purpose. To solve this problem, we used a 3D U-ResNet with pre- and postprocessing and data augmentation. The preprocessing includes the overlap-tile strategy by preparing the input patches, while a rule-based postprocessing was applied to remove false-positive artefacts. Our model achieved 0.812 average dice, 0.694 average surface dice and 0.7 tumor dice. This led to the 12.5th position in the KiTS21 challenge.

KiTS21 Challenge is to develop the best system for automatic semantic segmentation of renal tumors and surrounding anatomy. The organizers provide a dataset of 300 cases and each case’s CT scan is segmented to three semantic classes: Kidney, Tumor and Cyst. Compared with KiTS19 Challenge, cyst is a new semantic class, but these two tasks are quite close and that is why we choose nnUNet as our model and made some adjustments on it. Some important changes are made to the original nnUNet to adapt to this new task. Furthermore, we train models in 3 different ways and finally and merge them into one model by specific strategies. Detailed information is available in the part of Methods. The organizer uses an evaluation method called “Hierarchical Evaluation Classes” (HECs). The HEC scores of each model are showed in the following .

Accurate segmentation of kidney tumours can help doctors diagnose the disease. In this work, we described a multi-stage 2.5D semantic segmentation networks to automatically segment kidney and tumor and cyst in computed tomography (CT) images. First, the kidney is pre-segmented by the first stage network ResSENormUnet; then, the kidney and the tumor and cyst are fine-segmented by the second stage network DenseTransUnet, and finally, a post-processing operation based on a 3D connected region is used for the removal of false-positive segmentation results. We evaluate this approach in the KiTS21 challenge, which shows promising performance.

In this paper, we have described an automated algorithm for accurate segmentation of kidney, kidney tumors, and kidney cysts from CT scans. The Dataset for this problem was made available online as part of KiTS21 Challenge. Our approach was placed 13th in the official leaderboard of the competition. Our model uses a 2 stage Residual Unet architecture. The first network is designed to predict (Kidney + Tumor + Cyst) regions. The second network predicts segmented tumor and cyst regions from the output of the first network. The paper contains implementation details along with results on the official test and internal set.

Medical image processing plays an increasingly important role in clinical diagnosis and treatment. Using the results of kidney CT image segmentation for three-dimensional reconstruction is an intuitive and accurate method for diagnosis. In this paper, we propose a three-step automatic segmentation method for kidney, tumors and cysts, including roughly segmenting the kidney and tumor from low-resolution CT, locating each kidney and fine segmenting the kidney, and finally extracting the tumor and cyst from the segmented kidney. The results show that the average dice of our method for kidney, tumor and cysts is about 0.93, 0.57, 0.73.

As the most successful network structure in biomedical image segmentations, U-Net has presented excellent performance in many medical image segmentation tasks. We argue that the skip connections between the encoder and decoder layers pass too many redundant information, and filtered out the unnecessary information may be helpful in improving the segmentation accuracy. In this paper, we proposed a contrast attention mechanism at the skip connection, and proposed a contrast attention U-Net for KiTS21 challenge. The proposed method is able to achieve better performance over nnU-Net models with both low resolution and full resolution inputs in the 5-fold cross validation in the training set. In the final unrevealed testing set, our method achieves a mean sampled average Dice coefficient of 0.8815 and a mean sampled average surface Dice coefficient of 0.8007, which ranked the 5-th in the KiTS21 challenge.

Kidney cancer is one of the most common malignant tumors in the world. Automatic segmentation of kidney, kidney tumor, and kidney cyst is a essential tool for kidney cancer surgery. In this paper, we use a coarse-to-fine framework which is based on the nnU-Net and achieve accurate and fast segmentation of the kidney and kidney mass. The average Dice and surface Dice of segmentation predicted by our method on the test are 0.9077 and 0.8262, respectively. Our method outperformed all other teams and achieved \(1^{st}\) in the KITS2021 challenge.

Automatic segmentation of kidney tumors and lesions in medical images is an essential measure for clinical treatment and diagnosis. In this work, we proposed a two-stage cascade network to segment three hierarchical regions: kidney, kidney tumor and cyst from CT scans. The cascade is designed to decompose the four-class segmentation problem into two segmentation subtasks. The kidney is obtained in the first stage using a modified 3D U-Net called Kidney-Net. In the second stage, we designed a fine segmentation model, which named Masses-Net to segment kidney tumor and cyst based on the kidney which obtained in the first stage. A multi-dimension feature (MDF) module is utilized to learn more spatial and contextual information. The convolutional block attention module (CBAM) also introduced to focus on the important feature. Moreover, we adopted a deep supervision mechanism for regularizing segmentation accuracy and feature learning in the decoding part. Experiments with KiTS2021 testset show that our proposed method achieve Dice, Surface Dice and Tumor Dice of 0.650, 0.518 and 0.478, respectively.

Kidney cancer is one of the top ten cancers in the world, and its incidence is still increasing. Early detection and accurate treatment are the most effective control methods. The precise and automatic segmentation of kidney tumors in computed tomography (CT) is an important prerequisite for medical methods such as pathological localization and radiotherapy planning, However, due to the large differences in the shape, size, and location of kidney tumors, the accurate and automatic segmentation of kidney tumors still encounter great challenges. Recently, U-Net and its variants have been adopted to solve medical image segmentation problems. Although these methods achieved favorable performance, the long-range dependencies of feature maps learned by convolutional neural network (CNN) are overlooked, which leaves room for further improvement. In this paper, we propose an squeeze-and-excitation encoder-decoder network, named SeResUNet, for kidney and kidney tumor segmentation. SeResUNet is an U-Net-like architecture. The encoder of SeResUNet contains a SeResNet to learns high-level semantic features and model the long-range dependencies among different channels of the learned feature maps. The decoder is the same as the vanilla U-Net. The encoder and decoder are connected by the skip connections for feature concatenation. We used the kidney and kidney tumor segmentation 2021 dataset to evaluate the proposed method. The dice, surface dice and tumor dice score of SeResUNet are 67.2%, 54.4%, 54.5%, respectively.

Kidney cancer is aggressive cancer that accounts for a large proportion of adult malignancies. Computed tomography (CT) imaging is an effective tool for kidney cancer diagnosis. Automatic and accurate kidney and kidney tumor segmentation in CT scans is crucial for treatment and surgery planning. However, kidney tumors and cysts have various morphologies, with blurred edges and unpredictable positions. Therefore, precise segmentation of tumors and cysts faces a huge challenge. Consider these difficulties, we propose a cascaded deep neural network, which first accurately locate the kidney area through 2D U-Net, and then segment kidneys, kidney tumors, renal cysts through Multi-decoding Segmentation Network (MDS-Net) from the ROI of the kidney. We evaluated our method on the 2021 Kidney and Kidney Tumor Segmentation Challenge (KiTS21) dataset. The method achieved Dice score, Surface Dice and Tumor Dice of 69.4%, 56.9% and 51.9% respectively, in the test cases. The model of cascade network proposed in this paper has a promising application prospect in kidney cancer diagnosis.

Abdominal computed tomography is frequently used to non-invasively map local conditions and to detect any benign or malign masses. However, ill-defined borders of malign objects, fuzzy texture, and time pressure in fact, make accurate segmentation in clinical settings a challenging task. In this paper, we propose a two-stage deep learning architecture for kidney and kidney masses segmentation, denoted as convolutional computer tomography network (CCTNet). The first stage locates volume bounding box containing both kidneys. The second stage performs the segmentation of kidney, kidney tumors and cysts. In the first stage, we use a pre-trained 3D low resolution nnU-Net. In the second stage, we employ a mixup augmentation to improve segmentation performance of the second 3D full resolution nnU-Net. The obtained results indicate that CCTNet can provide improved segmentation of kidney, kidney tumor and cyst.

With the KiTS21 Grand Challenge, I propose the automatic segmentation model between the kidney and the mass of the kidney which includes tumor and cyst. Convolutional Neural Network is trained in patches of three-dimensional abdominal CT imaging. For the segmentation of the 3D image, a variant of U-Net which consists of 3D Encoder-Decoder CNN architecture with additional Skip Connection is used. Lastly, there is a loss function to resolve the class imbalance problem frequently occurring in the task of medical imaging. Sørensen-Dice Score and Surface Dice Score on the test set are 80.13 and 68.61.

Automatic segmentation of renal tumors and surrounding anatomy in computed tomography (CT) scans is a promising tool for assisting radiologists and surgeons in their efforts to study these scans and improve the prospect of treating kidney cancer. We describe our approach, which we used to compete in the 2021 Kidney and Kidney Tumor Segmentation (KiTS21) challenge. Our approach is based on the successful 3D U-Net architecture with our added innovations, including the use of transfer learning, an unsupervised regularized loss, custom postprocessing, and multi-annotator ground truth that mimics the evaluation protocol. Our submission has reached the 2nd place in the KiTS21 challenge.

Automated detection and segmentation of kidneys, tumors, and cysts are useful for renal diagnosis and treatment planning. Here we propose a two-stage contrast-enhanced CT detection and segmentation framework that automatically segments the kidney, kidney tumor, and cyst. Testing the proposed algorithm on the KiTS21 dataset, we achieve the mean dice of 0.5905 and the mean surface dice of 0.4234.

In order to compete in the KiTS21 challenge, we propose a 3D deep learning cascaded model for the renal enhanced CT image segmentation. The proposed model comprises two stages, where stage 1 segments the kidney and stage 2 segments the tumor and cyst. The proposed deep learning network architecture is based on the residual and 3D UNet architecture. The designed network is utilized for each segmentation stage (for stage 1 and stage 2). Our intended cascaded model achieved a dice score of 0.96 for the kidney, 0.81 for the tumor, and 0.45 for the cyst on the KiTS21 validation dataset.

This paper assesses whether using clinical characteristics in addition to imaging can improve automated segmentation of kidney cancer on contrast-enhanced computed tomography (CT). A total of 300 kidney cancer patients with contrast-enhanced CT scans and clinical characteristics were included. A baseline segmentation of the kidney cancer was performed using a 3D U-Net. Input to the U-Net were the contrast-enhanced CT images, output were segmentations of kidney, kidney tumors, and kidney cysts. A cognizant sampling strategy was used to leverage clinical characteristics for improved segmentation. To this end, a Least Absolute Shrinkage and Selection Operator (LASSO) was used. Segmentations were evaluated using Dice and Surface Dice. Improvement in segmentation was assessed using Wilcoxon signed rank test. The baseline 3D U-Net showed a segmentation performance of 0.90 for kidney and kidney masses, i.e., kidney, tumor, and cyst, 0.29 for kidney masses, and 0.28 for kidney tumor, while the 3D U-Net trained with cognizant sampling enhanced the segmentation performance and reached Dice scores of 0.90, 0.39, and 0.38 respectively. To conclude, the cognizant sampling strategy leveraging the clinical characteristics significantly improved kidney cancer segmentation. The model was submitted to the 2021 Kidney and Kidney Tumor Segmentation challenge.

The number of kidney cancer patients is increasing each year. Computed Tomography (CT) scans of the kidneys are useful to assess tumors and study tumor morphology. Semantic segmentation techniques enable the identification of kidney and surrounding anatomy on the pixel level. This allows clinicians to provide accurate treatment plans and improve efficiency. The large size of CT volumes poses challenges for deep segmentation methods as it cannot be accommodated on a single GPU in its original resolution. Downsampling CT scans influences the segmentation performance. In this paper, we present a coarse-to-fine cascaded network based on 3D U-Net architecture for semantic segmentation of kidney CT volumes into three classes kidney, tumor, and cyst. A two stage approach is implemented where a 3D U-Net model is first trained on downsampled CT volumes to delineate kidney region. This is followed by another 3D U-Net model which is trained using the full resolution images cropped around the areas of interest generated by first stage segmentation results. A set of 300 CT scans were used for training and evaluation. The proposed approach scored 0.9748, 0.8813, 0.8710 average dice for kidney, tumor and cyst using 3D cascade U-Net model. The performance of the cascade network outperformed other trained U-Net models based on 2D, 3D low resolution and 3D full resolution. The model also achieved the \(3^{rd}\) place in the leaderboard of KiTS21 challenge with a mean sampled average dice score of 0.8944 and a mean sampled average surface dice score of 0.8140 using a test set of 100 CT scans.

The accurate, automated detection and segmentation of renal tumors is of great interest for the imaging-based diagnosis, histologic subtyping, and management of suspected renal malignancy. The KiTS21 Grand Challenge provides 300 contrast enhanced CT images with kidney, tumors and cysts with corresponding manual annotation, to facilitate the development of robust segmentation algorithms for this task. In this work, we present an adaptation of the historically-successful 3D U-Net architecture, combined with deep supervision, foreground oversampling and large-scale image context, and trained on the majority-prediction segmentation masks. Our model achieved test-set performance of 97.0%, 85.1%, and 81.9% volumetric Dice score, and 93.7%, 72.0%, and 70.0% surface Dice score, on combined foreground, renal masses, and renal tumors, respectively, which tied for sixth place among challenge participants.

Kidney and Kidney tumor segmentation from CT scans has tremendous potential to help doctors in early diagnosis and localization of tumor, its size and type and for making timely treatment plans. However, considering the nature and volume of data, it is difficult and time consuming to train on such CT scans. In this paper, we propose enhancements to the 3D U-Net model to incorporate Spatial and Channel Attention in order to improve the identification and localization of segmentation structures by learning on spatial context. When compared with Residual U-Net model with greater depth and more feature maps, our Spatial and Channel Attention enhanced U-Net with less depth and feature maps performed significantly better on validation and training set when trained under similar conditions.

Transfer learning has witnessed a recent surge of interest after proving successful in multiple applications. However, it highly relies on the quantity of annotated data. Constrained by the labor cost and expertise, it is hard to annotate sufficient organs and tumors at the voxel level for medical image segmentation. Consequently, most bench-mark datasets were collected for the segmentation of only one type of organ and/or tumor, and all task-irrelevant organs and tumors were annotated as the background. We aim to make use of these partially but plentifully labeled datasets to boost the segmentation performance of the annotation-limited KiTS21 segmentation task. To this end, we first construct a general medical image segmentation model that learns to segment these partially labeled organs or tumors. Then we transfer its pre-trained weights to a specific downstream task, i.e., KiTS21. The primary experiments demonstrate the effectiveness of the proposed transfer learning strategy. Our method achieves 0.890 Dice score, 0.805 SurfaceDice, and 0.822 Tumor Dice in the KiTS21 challenge.