[Paper Review] Neural Network-Based Automatic Liver Tumor Segmentation With Random Forest-Based Candidate Filtering

Grzegorz Chlebus, Hans Meine+2 more|arXiv (Cornell University)|Jun 2, 2017

Brain Tumor Detection and Classification4 references59 citations

TL;DR

A fully automatic liver tumor segmentation method using cascaded CNNs for liver and tumor masks, followed by random forest-based filtering of tumor candidates to improve specificity.

ABSTRACT

We present a fully automatic method employing convolutional neural networks based on the 2D U-net architecture and random forest classifier to solve the automatic liver lesion segmentation problem of the ISBI 2017 Liver Tumor Segmentation Challenge (LiTS). In order to constrain the ROI in which the tumors could be located, a liver segmentation is performed first. For the organ segmentation, an ensemble of convolutional networks is trained to segment a liver using a set of 179 liver CT datasets from liver surgery planning. Inside of the liver ROI a neural network, trained using 127 challenge training datasets, identifies tumor candidates, which are subsequently filtered with a random forest classifier yielding the final tumor segmentation. The evaluation on the 70 challenge test cases resulted in a mean Dice coefficient of 0.65, ranking our method in the second place.

Motivation & Objective

• Motivate automatic liver lesion segmentation to reduce user interaction.
• Constrain tumor search to the liver ROI via robust liver segmentation.
• Identify tumor candidates with a CNN and filter false positives with a random forest.
• Achieve competitive LiTS challenge performance and analyze processing times.
• Discuss limitations and potential directions for 3D architectures and multi-label approaches.

Proposed method

• Ensemble of three orthogonal 2D U-nets (axial, sagittal, coronal) for initial liver segmentation.
• A 3D U-net refines and fuses the three 2D outputs into a final liver mask.
• A 2D U-net trained on liver voxels identifies tumor candidates within the liver ROI.
• A random forest classifier filters tumor candidates using 46 features from raw and refined masks.
• Postprocessing includes extracting 3D connected components and masking by the liver segmentation.

Experimental results

Research questions

• RQ1 Can cascaded CNNs reliably segment the liver and tumor regions in CT volumes for fully automatic LiTS submissions?
• RQ2 Does random forest-based candidate filtering improve specificity without sacrificing sensitivity in tumor segmentation?
• RQ3 What are the practical runtimes and limitations of the proposed pipeline on LiTS data?

Key findings

• Mean Dice coefficient on 70 LiTS test cases: 0.65 (second place).
• Total per-volume processing time: ~195 seconds (liver segmentation ~65 s, tumor segmentation ~49 s, postprocessing ~81 s).
• Tumor candidate classification accuracy with the random forest: 90%.
• Tumor segmentation is sometimes challenged by border tumors or inaccuracies in the liver mask, affecting final tumor masks.
• Training constrained to liver masks reduces training time but may induce false positives outside the liver region.
• Boundary patch sampling improves sensitivity at the cost of specificity.