Kidney and tumor segmentation using an ensemble of deep neural networks

Yu Wu, Yu Gan, Zhang Yi2019 · DOI

Abstract

For the segmentation of kidney and tumor task, we propose a two stages model that consists of several classification networks and segmentation models. The first stage is the foreground and background classification subnetwork, this stage is to recognize whether there are kidneys or tumors on images, so we propose a classification model called RD-Net which can effectively reduce the errors caused by a large of background images and improve the efficiency of the whole segmentation results. The second stage is the segmentation model used to predict the contour of the target (kidney or tumor). Therefore, we propose Att-ResUnet model and multi-scale ensemble of postprocessing methods used to integrate the predicted results of multiple models, so as to improve the accuracy of prediction results.

Introduction

Medical image segmentation, particularly for organs like kidneys and associated tumors, faces two major hurdles. First, a typical CT volume contains a vast number of "empty" slices where the target organ isn't present; feeding all these slices into a complex segmentation model is computationally wasteful and increases the risk of false positives. Second, the high variability in tumor shape and size makes it difficult for standard architectures to produce precise contours.

The authors tackle these issues by decoupling the task: they first identify if a target exists in a slice (classification) before attempting to determine where it is (segmentation).

Methods

The authors propose a two-stage framework consisting of a classification subnetwork followed by a segmentation pipeline.

Stage 1: The Classification Subnetwork (RD-Net)

To filter out background noise, the authors developed RD-Net. This model combines the strengths of two heavyweights: ResNet (known for residual connections that prevent vanishing gradients) and DenseNet (known for dense feature reuse).

Architecture of the RD-Net classification subnetwork.

Instead of choosing one, they concatenate the final feature maps of ResNet152 and DenseNet121 into a single 1536-dimensional vector before passing it to a fully-connected layer. Additionally, they use a "belief" algorithm for sequential post-processing. If the model misses a kidney in one slice but detects it in the slices immediately before and after, the algorithm corrects the middle slice to "1" (kidney present), leveraging the physical continuity of human anatomy.

Stage 2: The Segmentation Task

Once the relevant slices are identified, they are passed to a suite of segmentation models, including DeepLabv3, DeepLabv3+, and a custom Att-ResUnet.

Architecture of the Att-ResUnet segmentation model featuring attention gates.

The Att-ResUnet architecture replaces standard U-Net modules with residual blocks to allow for deeper learning. Crucially, it adds an attention mechanism to the upsampling (decoder) path. This helps the model focus on relevant features while suppressing irrelevant background noise. To ensure robustness, the authors employ a multi-scale ensemble where test images are resized into 10 different scales. The top 5 performing scales are then combined using a pixel-wise consensus voting method to generate the final segmentation mask.

Key Insights and Results

  • Hybridization and Context: The RD-Net hybrid outperformed individual ResNet and DenseNet models, proving that propagating errors back through both architectures simultaneously leads to better feature learning. Furthermore, treating the CT scan as a sequence rather than independent images significantly reduces classification errors.
  • Efficiency and Precision: By using the first-stage classifier to screen out background images, the system avoids the computational cost of running heavy segmentation models on irrelevant data. The attention gates in Att-ResUnet act as a filter, allowing the model to "ignore" the majority of the image and focus on the boundaries of the kidney/tumor.
  • Ensemble Stability: The multi-scale ensemble helped stabilize the final output, ensuring that the predicted contours were not artifacts of a specific image size.

Limitations and Future Work

While the paper demonstrates high performance, the ensemble approach is computationally expensive for real-time clinical use. Future work might explore more tightly integrated "one-stage" architectures that can handle classification and segmentation simultaneously without losing the benefits of background filtering. Additionally, further research is needed to determine how well this model generalizes to other organs or different CT hardware.

Summary updated

Abstract

For the segmentation of kidney and tumor task, we propose a two stages model that consists of several classification networks and segmentation models. The first stage is the foreground and background classification subnetwork, this stage is to recognize whether there are kidneys or tumors on images, so we propose a classification model called RD-Net which can effectively reduce the errors caused by a large of background images and improve the efficiency of the whole segmentation results. The second stage is the segmentation model used to predict the contour of the target (kidney or tumor). Therefore, we propose Att-ResUnet model and multi-scale ensemble of postprocessing methods used to integrate the predicted results of multiple models, so as to improve the accuracy of prediction results.
BibTeX
@misc{wu2019kidney,
  author = {Yu Wu and Yu Gan and Yuhang Wu and Zhang Yi},
  title = {{Kidney and tumor segmentation using an ensemble of deep neural networks}},
  year = {2019},
  doi = {10.24926/548719.026},
}
Figure