[Paper Review] 2D-Densely Connected Convolution Neural Networks for automatic Liver and Tumor Segmentation

Krishna Chaitanya Kaluva, Mahendra Khened+2 more|arXiv (Cornell University)|Jan 5, 2018

Advanced Neural Network Applications6 references68 citations

TL;DR

The paper proposes a two-stage cascaded DenseNet-based framework for automatic liver and tumor segmentation in CT images, achieving liver dice of 0.923 and tumor dice of 0.625 on the test set. It uses liver localization to guide tumor segmentation within the liver region and employs weighted losses and post-processing to improve boundary accuracy.

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 $(256 imes 256)$, 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.

Motivation & Objective

• Develop an automatic method to segment the liver and its tumors from CT volumes.
• Leverage a two-stage cascaded approach where liver segmentation localizes the tumor segmentation region.
• Use densely connected fully convolutional networks to enhance feature reuse with limited data.

Proposed method

• Adopt a two-stage cascaded framework: first segment the liver, then segment tumors within the liver region.
• Employ DenseNet-based FCN blocks with 4-layer dense blocks and transition layers for down-sampling and up-sampling.
• Train the liver model on down-sampled 256x256 axial slices; train tumor model on full-size slices with three different HU windowings as input.
• Use spatially weighted cross-entropy loss with edge-weight maps to improve contour accuracy for the liver; employ a mixed loss for tumors including dice loss.
• Apply post-processing (erosion, largest connected component, dilation) to liver predictions to reduce false positives.

Experimental results

Research questions

• RQ1 Can a cascaded DenseNet-based FCN accurately segment liver and tumors in CT data using stage-wise localization?
• RQ2 Does multi-window HU input improve tumor boundary learning compared to single-channel input?
• RQ3 What loss formulation best handles class imbalance and boundary accuracy for liver and tumor segmentation?
• RQ4 What post-processing steps optimally reduce false positives in liver segmentation?
• RQ5 How does the proposed method perform on the LITS MICCAI-2017 dataset?

Key findings

• Liver segmentation achieved an average dice score of 0.923 on the 14-test-volume set.
• Tumor segmentation achieved an average dice score of 0.625 on the same test set.
• Tumor burden estimation yielded RMSE of 0.044 on the test set.
• The proposed 2-stage cascade with liver-guided tumor localization improved segmentation within liver regions.
• Use of 3-channel HU-windowed inputs for tumors improved boundary learning compared to a single-channel approach.
• Spatially weighted cross-entropy combined with dice loss provided better overlap than cross-entropy alone.