2D-Densely Connected Convolution Neural Networks for automatic Liver and Tumor Segmentation

Krishna Chaitanya Kaluva, Mahendra Khened, Avinash Kori, Ganapathy KrishnamurthiarXiv 2018 · OpenReview

TL;DR

This paper introduces a two-stage, fully automatic system for segmenting livers and tumors in CT scans using DenseNet. By first segmenting the liver and then using that output to guide tumor segmentation, they achieve impressive accuracy, paving the way for more efficient and reliable medical image analysis.

What Problem Does This Paper Address?

Manually segmenting organs like the liver and identifying tumors within them in CT scans is a time-consuming and labor-intensive process, prone to inter-rater variability. While deep learning, particularly Convolutional Neural Networks (CNNs), has shown promise in medical image segmentation, accurately segmenting both the liver and its often diffusely located tumors, especially in the presence of class imbalance (tumors being a small fraction of the image), remains a challenge. This paper aims to develop an automated, accurate, and efficient method for this dual segmentation task.

Methods

The core of the proposed solution is a two-stage cascaded approach leveraging DenseNet, a CNN architecture known for its efficient feature reuse.

  • Stage 1: Liver Segmentation: A DenseNet-based fully convolutional network (FCN) is trained to segment the liver. To improve efficiency and reduce computational load, the CT slices are down-sampled to 256x256 pixels. The network uses a specific architecture with dense blocks, transition down layers, and a bottleneck. A key innovation here is the use of "spatially-weighted-cross-entropy loss" combined with L2 regularization. This loss function assigns higher penalties to misclassifications near the liver's edges, ensuring more precise contouring.
  • Stage 2: Tumor Segmentation: A separate DenseNet-based FCN is trained for tumor segmentation. This model takes a 3-channel input, where each channel is the same CT slice but windowed at different Hounsfield Unit (HU) ranges. This multi-channel input helps the model better capture tumor characteristics. Unlike the liver model, this stage does not down-sample the slices. The loss function here is a weighted combination of spatially-weighted-cross-entropy loss and a Dice loss, aiming to address class imbalance and improve overlap accuracy.
  • Cascading: The output of the liver segmentation model from Stage 1 is used to mask the input for Stage 2. This ensures that tumor segmentation is localized within the liver region, preventing false positives from surrounding tissues.
  • Post-processing: Both stages employ post-processing steps. For the liver, this includes morphological operations like erosion and dilation to refine the segmentation and remove small false positives. For tumors, the prediction is masked by the liver segmentation.

Datasets

The models were trained and evaluated on the LITS MICCAI-2017 challenge dataset, which consists of 200 contrast-enhanced CT scans. The dataset was split into training (90 volumes), validation (26 volumes), and testing (14 volumes) sets for the liver model. For the tumor model, the split was train: 90 volumes, validation: 26 volumes, and test: 14 volumes. The paper also mentions using 70 CT volumes for testing the combined model's segmentation metrics.

Key Insights and Ideas

  • Cascaded Approach for Dual Segmentation: The two-stage cascade (liver first, then tumor guided by liver) is a crucial insight. It effectively breaks down a complex problem into manageable steps, leveraging the robust segmentation of a large organ to aid in the detection of smaller, more challenging lesions.
  • Spatially-Weighted Loss for Precise Contours: The use of spatially-weighted loss functions, particularly for liver segmentation, is a clever way to enforce accurate boundary detection, which is often a weak point in standard segmentation models.
  • Multi-Channel Input for Tumor Detection: Feeding the tumor segmentation network with CT slices windowed at different HU ranges significantly improves its ability to discern tumors from surrounding tissues, addressing the challenge of diffuse and inhomogeneous tumor appearances.
  • DenseNet's Efficiency: The adoption of DenseNet highlights its effectiveness in medical imaging tasks, where data can sometimes be limited. Its dense connectivity allows for better feature reuse and reduces the number of parameters, making it more efficient.

Key Results

The proposed cascaded DenseNet model achieved impressive results on the test set:

  • Liver Segmentation: A global Dice score of 0.923.
  • Tumor Segmentation: A global Dice score of 0.625.
  • Tumor Burden Estimation: An RMSE of 0.044.

The paper also notes that the model produces smoother liver surfaces compared to U-Net, which can exhibit "ridges" due to its skip connections.

Limitations and Future Work

  • Tumor Detection False Positives: The paper acknowledges that the model sometimes misclassifies darker vessels as tumors.
  • Tumor Segmentation Accuracy: While improved, the tumor segmentation Dice score (0.625) is still lower than that for the liver, indicating room for further improvement in capturing the intricate details of tumors.
  • Future Work: The authors suggest that further improvements could be made by exploring more advanced post-processing techniques and potentially incorporating 3D convolutional networks for a more holistic volumetric analysis.

Summary updated

Abstract

In this paper we propose a fully automatic 2-stage cascaded approach for segmentation of liver and its tumors in CT (Computed Tomography) images using densely connected fully convolutional neural network (DenseNet). We independently train liver and tumor segmentation models and cascade them for a combined segmentation of the liver and its tumor. The first stage involves segmentation of liver and the second stage uses the first stage's segmentation results for localization of liver and henceforth tumor segmentations inside liver region. The liver model was trained on the down-sampled axial slices $(256 \times 256)$, whereas for the tumor model no down-sampling of slices was done, but instead it was trained on the CT axial slices windowed at three different Hounsfield (HU) levels. On the test set our model achieved a global dice score of 0.923 and 0.625 on liver and tumor respectively. The computed tumor burden had an rmse of 0.044.
BibTeX
@misc{kaluva2018d,
  author = {Krishna Chaitanya Kaluva and Mahendra Khened and Avinash Kori and Ganapathy Krishnamurthi},
  title = {{2D-Densely Connected Convolution Neural Networks for automatic Liver and Tumor Segmentation}},
  year = {2018},
  eprint = {1802.02182},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/1802.02182},
}
Figure