Source-linked AI summary
Automatic Liver Lesion Segmentation Using A Deep Convolutional Neural Network Method
Xiao Han
TL;DR
Liver lesion segmentation is clinically important but difficult because lesions vary widely and CT images have limited soft-tissue contrast. The paper develops a 2.5D DCNN combining U-Net and ResNet connections, achieving an average Dice value of 0.67 on LiTS test data. The authors conclude that accuracy remains rather low and requires further improvement.
Problem
Liver lesion segmentation is challenging because lesions vary across patients and CT images have low soft-tissue contrast and artifacts.
Method
The paper develops a 2.5D DCNN using adjacent axial slices with U-Net long-range and ResNet short-range connections.
Results
0.67 average Dice value was achieved on the LiTS test data.
Takeaways & Limitations
The DCNN framework supports end-to-end training using largely original image data without manually designed image features or preprocessing.
Takeaways & Limitations
The authors state that lesion segmentation accuracy remains rather low and that further improvements are needed.
Abstract
from arXiv · showhide
Liver lesion segmentation is an important step for liver cancer diagnosis, treatment planning and treatment evaluation. LiTS (Liver Tumor Segmentation Challenge) provides a common testbed for comparing different automatic liver lesion segmentation methods. We participate in this challenge by developing a deep convolutional neural network (DCNN) method. The particular DCNN model works in 2.5D in that it takes a stack of adjacent slices as input and produces the segmentation map corresponding to the center slice. The model has 32 layers in total and makes use of both long range concatenation connections of U-Net [1] and short-range residual connections from ResNet [2]. The model was trained using the 130 LiTS training datasets and achieved an average Dice score of 0.67 when evaluated on the 70 test CT scans, which ranked first for the LiTS challenge at the time of the ISBI 2017 conference.
1. INTRODUCTION
Liver lesion segmentation supports diagnosis, treatment planning, and treatment evaluation, but manual 3D CT annotation is slow and variable. The paper addresses challenging lesion variability and limited prior success with a 2.5D DCNN combining U-Net and ResNet design elements.
- Liver lesion segmentation can support cancer diagnosis, treatment selection, and evaluation of treatment effectiveness.
- Manual segmentation of 3D CT images is time-consuming and subject to inter- and intra-rater variation.
- Lesions vary substantially in location, size, shape, intensity, texture, and occurrence count across patients, while CT images have low soft-tissue contrast and artifacts.
- Existing intensity-clustering, region-growing, and deformable-model methods have shown limited success for automatic lesion segmentation.
- The proposed DCNN uses adjacent axial slices as input and combines U-Net and ResNet architectural features in a 2.5D design.
2. DATASET AND PREPROCESSING
The study uses heterogeneous LiTS abdominal CT scans from multiple clinical sites, scanners, and protocols. Preprocessing is minimal, while two models use coarse resampling for rapid liver localization and original resolution for detailed lesion segmentation.
- The LiTS datasets contain 200 contrast-enhanced 3D abdominal CT scans from multiple clinical sites, scanners, and protocols.
- In-plane resolution ranges from 0.60 mm to 0.98 mm, and slice spacing ranges from 0.45 mm to 5.0 mm.
- All axial slices are 512 × 512, while scan slice counts vary from 42 to 1026.
- Image intensities were truncated to [−200, 200] HU, with no other special preprocessing described.
- Two DCNN models reduce computation: a coarse-resolution model localizes the liver, while an original-resolution model performs detailed liver and lesion segmentation.
3. METHOD
The method uses a fully convolutional, 2.5D DCNN that combines U-Net and ResNet connections for end-to-end liver and lesion segmentation. It uses a two-model pipeline, weighted training, and LiTS-specific preprocessing and deployment procedures.
- DCNN architecture: The fully convolutional network segments an entire 2D image in one pass, using broader contextual information than patch-based classification.This design is intended to improve efficiency and segmentation reliability.
- DCNN architecture: The model combines long-range U-Net copying and concatenation connections with short-range ResNet residual connections.The residual connections support information propagation through the network, while U-Net connections provide high-resolution features to decoding layers.
- 2.5D input and output: Five adjacent axial slices form the input, and the network outputs a segmentation map for the center slice.This 2.5D design provides axial content and orthogonal contextual information while using 2D output maps.
- Training objective: Weighted cross-entropy trains the end-to-end mapping, with weights 0.2 for background, 1.2 for liver, and 2.2 for liver lesion.The weighting factors are class-dependent and are specified empirically.
- Training and inference pipeline: Two models are trained: a coarse liver model on resampled scans and a detailed model using original-resolution slices and background, liver, and lesion labels.The second model processes only slices inside the liver region to support detailed lesion segmentation.
- Implementation: Training uses random 320×320×5 crops, random left-right flips, batch normalization, and stochastic gradient descent with momentum for 50 epochs.Each model required about 4 days to train on a single NVIDIA Titan X GPU; inference took about 0.2 second per slice.
4. POSTPROCESSING
Postprocessing removes likely false-positive lesion components after connected-component analysis. The submitted LiTS results used an empirically set probability threshold and achieved an average Dice value of 0.67.
- Lesion-component filtering: A 3D connected-component analysis removes lesion components whose maximum DCNN lesion probability is below 0.80.The threshold was set empirically for the final LiTS competition results.
- Reported result: 0.67 average Dice was achieved on the LiTS validation data after postprocessing.The reported value is the overall average Dice value for the submitted method.
5. DISCUSSION
The DCNN framework supports end-to-end training from largely original image data, but network design remains difficult to optimize and lesion accuracy still requires improvement.
- End-to-end DCNN training uses largely original image data without manually designed features or preprocessing.
- DCNN architecture selection remains difficult because different designs matter, yet no clear application-specific optimum exists and evaluating alternatives is time-consuming.
- Lesion segmentation accuracy remains rather low, motivating investigation of improved architectures and postprocessing strategies such as 3D conditional random fields, cascades, or ensembles.