Fully Automatic Segmentation of Kidney and Tumor Based on Cascaded U-Nets

Yu Li2019 · DOI

Abstract

Fully automatic segmentation of the kidney and its associated lesions is a critical step toward achieving accurate clinical diagnoses and developing robust computer-aided decision support systems. This paper presents a method for the automatic segmentation of kidneys and renal tumors in abdominal CT images using a cascaded 3D U-Net convolutional neural network (3D cU-Net). By training and cascading two 3D U-Nets, the system first localizes the kidney to define a region of interest (ROI) and subsequently segments the lesion within that ROI. This hierarchical approach effectively addresses challenges such as low image contrast and extreme class imbalance.

Introduction

Segmenting kidney tumors is a high-stakes task in medical imaging, as accurate boundaries are essential for surgical planning and radiotherapy. Traditionally, this process is performed manually, which is time-consuming, subjective, and prone to human error. From a computational perspective, the task is hindered by three primary factors:

  1. Low Contrast: Kidney tumors often exhibit a grayscale appearance similar to surrounding healthy tissue.
  2. Morphological Irregularity: Tumors vary significantly in size, shape, and location across patients.
  3. Class Imbalance: In a full abdominal CT volume, the tumor occupies only a tiny fraction of the total data, making it difficult for standard models to distinguish small lesions from the vast background.

Methodology: The Cascaded 3D U-Net

To overcome these hurdles, the authors propose a 3D Cascaded U-Net (3D cU-Net) architecture. Instead of employing a single model for the entire task, the problem is decomposed into two specialized stages:

  • Step 1: Kidney Localization. A 3D U-Net is trained to segment the entire kidney. This output serves as a "Region of Interest" (ROI) selector, effectively cropping the volumetric data to focus exclusively on the organ.
  • Step 2: Tumor Segmentation. A second 3D U-Net takes the renal ROI predicted in the first step and focuses its parameters on identifying lesions within that specific volume.
Architecture of the 3D U-Net used in the cascaded framework.

Technical Implementation

The 3D architecture processes volumetric "voxels" rather than individual 2D slices, allowing the model to leverage spatial relationships across the Z-axis. To optimize training, a hybrid loss function combining Cross-Entropy and Dice Loss was utilized. Dice loss is particularly valuable in medical segmentation as it measures the overlap between predicted and ground-truth masks, preventing the model from becoming biased toward the background class.

Data and Preprocessing

The study utilized a dataset of 300 abdominal CT images from the University of Minnesota Medical Center, consisting of 210 training and 90 test images. The cohort included patients undergoing partial or radical nephrectomy, providing a diverse range of clinical scenarios.

Preprocessing involved Hounsfield Unit (HU) windowing (specifically [-160, 240]) to filter out irrelevant tissues like bone or air. This step enhances the contrast of soft tissues, making the kidney and potential tumors more distinguishable for the neural network.

Results and Discussion

The cascaded approach demonstrated high adaptability to the challenging task of renal lesion detection. By "zooming in" on the kidney, the second network avoids interference from other organs like the liver or spine, leading to more precise tumor boundaries.

Extracted figure from page 6 of Fully Automatic Segmentation of Kidney and Tumor Based on Cascaded U-Nets

Furthermore, the cascading method improves memory efficiency. Processing full-resolution 3D abdominal scans is computationally expensive; by cropping the volume after the first stage, the system maintains high resolution where it is most needed without exceeding GPU memory limits. The authors also implemented anatomical post-processing to remove false positives by verifying that predicted segments align with known human anatomy (e.g., the typical bilateral symmetry of kidneys).

Conclusion and Future Work

The 3D cU-Net provides an objective "second opinion" for clinicians, potentially reducing the reliance on invasive renal biopsies by providing high-fidelity diagnostic data from non-invasive CT scans. While qualitative results are promising, future work should focus on:

  • Quantitative Validation: Explicitly detailing Dice coefficients and Hausdorff distances across larger, multi-center test sets.
  • Generalization: Testing the model's robustness against different CT scanner brands and imaging protocols.
  • Radiomics Integration: Moving beyond segmentation to predict tumor pathology (benign vs. malignant) directly from the imaging data.

Summary updated

Abstract

Fully automatic segmentation of kidney and its lesions is an important step to obtain accurate clinical diagnosis and computer aided decision support system. In this paper, a method of automatic segmentation of kidney and renal tumor in CT abdominal images using cascade 3D U-Net convolutional neural network (3D cU-Nets) is presented. We trained and cascaded two 3D U-Nets for the joint segmentation of kidney and renal tumor. In the first step, we trained a 3D U-Net to segment kidney as the ROI input for the second 3D u-net.The second 3D U-Net only segmented the lesion from the renal ROI predicted in step 1.
BibTeX
@misc{li2019fully,
  author = {Yu Li},
  title = {{Fully Automatic Segmentation of Kidney and Tumor Based on Cascaded U-Nets}},
  year = {2019},
  doi = {10.24926/548719.008},
}
Figure