2019 Kidney Tumor Segmentation Challenge Method Manuscript

MengLei Jiao, Hong Liu2019 · DOI

TL;DR

This paper presents a two-stage deep learning framework for segmenting kidneys and kidney tumors in CT scans, achieving high accuracy on the KiTS19 challenge dataset. The approach combines a novel ResUNetSM model for kidney segmentation with a DeepLabV3+ model for tumor segmentation within a refined region of interest, enhanced by post-processing techniques.

What Problem Does This Paper Address?

Accurate segmentation of kidneys and kidney tumors from CT scans is crucial for diagnosis, treatment planning, and monitoring. However, this task is challenging due to significant variations in tumor size, shape, location, and intensity across patients. Existing methods, while improving, still struggle with precise delineation, especially in complex cases. The KiTS19 challenge aimed to drive progress in this area by providing a large dataset and a standardized evaluation framework.

Methods

The authors propose a two-stage segmentation pipeline:

  1. Kidney Segmentation (ResUNetSM): A novel 2D U-Net-like architecture, termed ResUNetSM, is introduced. It leverages a ResNet101 encoder for feature extraction and a decoder incorporating SELayer (for channel interdependencies) and MobileBlock (for efficiency) modules. The architecture also includes an Atrous Spatial Pyramid Pooling (ASPP) layer for multi-scale feature capture and skip connections between encoder and decoder for preserving fine details.
  2. Tumor Segmentation (DeepLabV3+): A 2.5D DeepLabV3+ model is employed for tumor segmentation. This model takes multiple adjacent axial slices as input. Crucially, tumor segmentation is performed only within a 3D Region of Interest (ROI) derived from the initial kidney segmentation. This strategy aims to reduce noise and focus the model on relevant areas.
  3. Post-processing: To further refine the segmentation, the authors apply two post-processing techniques:
  4. 3D Conditional Random Fields (3DCRF): This method smooths the segmentation boundaries and incorporates spatial context.
  5. 3D Connected Component Analysis (3DCCA): This technique removes small, spurious segmented regions and consolidates larger ones.

The entire pipeline combines the kidney and tumor segmentation results.

Datasets

The study utilizes the KiTS19 dataset, which comprises 300 abdominal CT scans. For this work, 210 scans were used for training and validation, with the remaining 90 reserved for testing. The authors split the 210 scans into a training set (195 scans) and a validation set (15 scans). The CT scans have varying slice thicknesses (1-5mm) and resolutions (512x512 pixels), with slice counts ranging from 29 to 1059 per scan. The dataset contains annotations for kidney and tumor regions.

Key Insights and Ideas

  • Two-Stage Segmentation for Focus: Performing tumor segmentation within a pre-defined kidney ROI significantly improves accuracy by reducing false positives and focusing computational resources. This is a smart way to leverage the output of the first stage to guide the second.
  • Hybrid Network Architecture (ResUNetSM): Combining established powerful components like ResNet with efficient modules like MobileBlock and attention mechanisms like SELayer in a U-Net structure offers a robust solution for kidney segmentation. The inclusion of ASPP and skip connections further enhances its capability.
  • Importance of Post-processing: The significant improvements observed after applying 3DCRF and 3DCCA highlight that even state-of-the-art deep learning models benefit greatly from traditional image processing techniques for refining segmentation masks, especially in medical imaging.
  • Oversampling for Imbalanced Data: The authors employed oversampling for the tumor class during training of DeepLabV3+, acknowledging the inherent class imbalance (tumors are often smaller and less frequent than kidneys) and its impact on model performance.

Key Results

The proposed framework achieved impressive results on the validation set:

  • Kidney Segmentation: A mean Dice score of 96.31%.
  • Tumor Segmentation: A mean Dice score of 81.64%.
  • The combined kidney-tumor segmentation achieved a mean Dice score of 88.97%.

These results were obtained after applying both 3DCRF and 3DCCA post-processing. The paper shows that post-processing significantly boosted the tumor segmentation Dice score from 0.7424 (base model) to 0.8164.

Limitations and Future Work

  • Validation Set Bias: The authors acknowledge that the results are based on the validation set and that there might be some bias. The true performance on the unseen test set will only be known after the leaderboard is announced.
  • Generalizability: While the model performed well on the KiTS19 dataset, its performance on other datasets with different CT scanner protocols or patient populations is not discussed.
  • Computational Cost: Although MobileBlocks were used for efficiency, the overall pipeline involves multiple deep learning models and post-processing steps, which can still be computationally intensive.

Future work could involve exploring more advanced 3D segmentation architectures, investigating methods to further reduce computational cost, and evaluating the model's robustness across diverse datasets.

Summary updated

Abstract

This paper framework in detail for KiTS19, which is the 2019 Kidney Tumor Segmentation Challenge. We adopt two model ResUNetSM and DeepLabV3 plus to segment kidney and tumor respectively. Firstly, we propose a model ResUNetSM to segment kidney, which uses ResNet for encoder, and adopts SELayer and MobileBlock for decoder. ResUNetSM also adopts ASPP and skip-connect structure. To segment tumor region, we adopt DeepLabV3 plus and segment tumor in the 3D ROI region from above kidney segmentation results to reduce noise. Finally, we use 3DCRF and 3D connected component analysis as post-processing to improve the final segmentation results. Our framework gets the 96.31% mean dice for kidney and 81.64% mean dice for tumor on validation set.
BibTeX
@misc{jiao2019kidney,
  author = {MengLei Jiao and Hong Liu},
  title = {{2019 Kidney Tumor Segmentation Challenge Method Manuscript}},
  year = {2019},
  doi = {10.24926/548719.042},
}
Figure