Cascaded Volumetric Convolutional Network for Kidney Tumor Segmentation from CT volumes

Yao Zhang, Yixin Wang, Feng Hou, Jiawei Yang, Guangwei Xiong, Jiang Tian, Cheng Zhong

Introduction

Kidney cancer is a significant global health concern, affecting over 400,000 people annually. Precise surgical planning and treatment depend heavily on accurate understanding of tumor morphology, typically derived from 3D Computed Tomography (CT) scans. However, automated segmentation of these volumes presents two major hurdles: extreme class imbalance and high computational costs.

  1. Class Imbalance: The kidney and its tumors occupy only a tiny fraction of the total abdominal CT volume, making it easy for standard models to lose focus amidst irrelevant background tissue.
  2. Computational Constraints: While 3D Fully Convolutional Networks (FCNs) capture spatial context better than 2D counterparts, their memory intensity often necessitates trade-offs in resolution or network depth.

To overcome these issues, this research introduces a two-stage cascaded 3D convolutional neural network designed for memory efficiency and high-precision segmentation.

Methodology: The Two-Stage Cascade

The core of the proposed approach is a two-stage cascaded framework that breaks the segmentation task into localization and refinement steps.

Overview of the two-stage cascaded framework for kidney and tumor segmentation.
  • Stage 1: Coarse Localization. A basic 3D U-Net processes a downsampled version of the CT volume to identify the kidney's general location. The image is then cropped around this region, drastically reducing the search space and background noise for the subsequent stage.
  • Stage 2: Fine Segmentation. A more sophisticated multitask 3D U-Net operates on the high-resolution cropped region to segment both the kidney and the tumor simultaneously.

Architectural Enhancements

The framework utilizes several advanced deep learning techniques to optimize performance:

  • Residual Blocks: These allow for better gradient flow through shortcut connections, enabling more effective learning in deep architectures.
  • Deep Supervision (DS): Loss is calculated at multiple levels of the decoder, forcing intermediate layers to learn more discriminative features.
  • Spatial Priors (SP): The kidney mask generated in Stage 1 is fed into Stage 2 as an additional input channel. This acts as a guide, focusing the second model's attention on the relevant anatomy.

Experimental Results

The model was evaluated using the MICCAI 2019 KiTS (Kidney Tumor Segmentation) Challenge dataset, which consists of 300 arterial-phase CT scans. Preprocessing involved percentile clipping to handle intensity outliers and standard normalization.

Ablation study showing the performance gains from 3D context and specific architectural enhancements.

As shown in the ablation study, each component contributed significantly to the final performance. The transition from a 2D U-Net baseline to the full 3D cascaded model with spatial priors saw the Tumor Dice score jump from 62.35% to 83.70%. The final system achieved a Kidney Dice score of 98.05% and a Composite Dice score of 90.24%, securing 4th place out of 105 teams in the KiTS 2019 Challenge.

Conclusion

By adopting a "zoom-in" strategy, the cascaded volumetric network effectively bypasses the memory limitations of 3D networks while addressing the class imbalance inherent in medical imaging. While the two-stage process adds complexity to the deployment pipeline, the substantial gains in segmentation accuracy—particularly for small tumors—demonstrate the power of combining global context with local refinement.

Summary updated

BibTeX
@misc{zhang2019cascaded,
  author = {Yao Zhang and Yixin Wang and Feng Hou and Jiawei Yang and Guangwei Xiong and Jiang Tian and Cheng Zhong},
  title = {{Cascaded Volumetric Convolutional Network for Kidney Tumor Segmentation from CT volumes}},
  year = {2019},
  eprint = {1910.02235},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/1910.02235},
}
Figure