Rencheng Zheng1, Qidong Wang2, Ziying Feng3, Chengyan Wang3, and He Wang1,3
1Institute of Science and Technology for Brain-Inspired Intelligence, Fudan University, Shanghai, China, 2Radiology department, The first affiliated hospital, College of medicine, Zhejiang University, Hangzhou, China, 3Human Phenome Institute, Fudan University, Shanghai, China
Synopsis
The objective of this study is to perform automatic multi DCE phases liver
tumor detection using deep learning based segmentation and radiomics guided candidate
filtering. The proposed model consists mainly of two stages, primary segmentation
based on a U-net architecture neural network in stage1, and suspected tumor discrimination
mechanism using multi DCE phases radiomics features including shape features,
texture features, time dimension features and location information in stage 2. The
proposed two-stage model exhibits superior performance in HCC tumor segmentation
with a mean Dice score of 0.7928 in test set.
INTRODUCTION
Liver cancer is one of the most common cause of cancer-induced deaths
according to the statistic from World Health Organization. Hepatocellular
carcinoma (HCC) as the most common type of primary liver cancer, which is the
sixth most prevalent cancer1. Early diagnosis and
treatment for HCC are crucial for improving the success rate of tumor resection
and survival rate2. Accurate tumor segmentation
is an important prerequisite for quantitative analysis and evaluation of tumors,
which can help doctors better understand tumors and develop appropriate
treatment plans. Magnetic Resonance (MR) Dynamic Contrast Enhanced (DCE)
imaging is a common imaging modality in liver cancer, this study proposed a
model combining multi DCE phases to address the liver tumor detection task
through deep learning based segmentation and radiomics features guided
candidate filtering.METHODS
95 HCC patients were recruited into the subject group for gadolinium-based
contrast agent enhanced MR scans. Four DCE phases (mask phase, arterial phase,
portal venous phase, delayed phase) were acquired for each participate. The
whole patient set was randomly divided into train set (71 cases, for neural network
and discriminator training), validation set (12 cases, for evaluation and hyperparameter
adjustment) and test set (12 cases, for generalization ability evaluation). The
liver tumor detection was mainly divided into two stages, the first stage is primary
segmentation, in which (a) a classical 2D U-net architecture neural network with
‘batch normalization’ was adopted for liver segmentation. The input of the
network was 4-channel co-registered DCE phases, where the registration was based
on Symmetric Normalization (SyN) algorithm3, performed in
Advanced Normalization Tools (ANTs). (b) Areas outside the liver were regarded
as background
and would not be
considered in subsequent tumor segmentation. (c) The segmented liver ROI was
used as the input of the second neural network for tumor segmentation. Tumor
candidates generated after a relatively small threshold (0.3 in practice) in tumor
segmentation probability map to ensure that most suspected tumor areas were
included in. The second stage was suspected tumor candidates filtering based on
radiomics features. (d) Each suspected tumor area was identified as a 3D connecting
component for radiomics features extraction, texture features and common shape
features were extracted in four DCE phases. Beyond that some time dimension
features were extracted as well, including the difference of texture features
on the timeline and time curve features (kurtosis, skewness, mean and variance).
Additionally, there were two other features describing the location of the
tumor candidate. In general, 1108 radiomics features were included in the
feature set. (e) The least absolute shrinkage and selection operator (LASSO)
logistic regression4 was applied to do
the feature selection, and the top ten essential elements were used in the classification
model. (f) Various classifiers were trained and evaluated in the 5-fold
cross-validation, the optimal classifier was chosen to predict in the test set.
After the two-stage segmentation and discrimination, the identified HCC tumor components
were regarded as the final results. 3D volume Dice similarity coefficient (Dice)
score was adopted to evaluate the segmentation performance. The whole framework
is shown in Figure 1.RESULTS
The
liver segmentation achieved mean Dice score of 0.9610 in test set. The primary
tumor segmentation based on multi DCE phases in the first stage achieved mean Dice
score of 0.7237, which was more dominant than tumor segmentation based on one
DCE phase (Figure 2 (a)). After the second
stage discrimination using radiomics features, the final mean Dice score
improved to 0.7928 (Figure 2 (b)). Compared to discrimination
mechanism without time dimension features, the proposed model had a better discrimination
performance, manifested in accuracy, area under curve (AUC) value, F1 score, false
positive rate (FPR) and final segmentation Dice scores (Figure 3). A segmentation slice
demonstration is illustrated in Figure 4.DISCUSSION
The results indicated that the proposed automatic two-stage segmentation
model has a superior performance in liver tumor segmentation compared to some
existed methods5, 6. High false positive rate is common for small region detection after primary
segmentation through neural network, hence a discrimination mechanism was
adopted to filter the ‘fake tumor’ for a better segmentation performance. Typically,
shape features were considered important
in discrimination. However, due to the relatively poor segmentation results in
test set compared to training set, the shape characteristics might be destroyed
and show a weakened influence
in discrimination. In this study, time dimension features in multi DCE phases
were extracted and considered to be
helpful for true tumors identification. The reason was that due to the
influence of contrast agent, the response of tumor region in multi DCE phases
was different from external areas. This difference could be reflected in the
time dimension features and generally not affected by the segmentation effect,
resulted in a better discrimination performance. CONCLUSION
This study demonstrated an automatic two-stage liver tumor segmentation
model, which consists of a primary segmentation model based on deep learning
and a discrimination mechanism guided by radiomics features. The proposed model
had a considerable performance in HCC tumor segmentation.Acknowledgements
Acknowledge: This work was supported by Shanghai Municipal Science and Technology Major Project (No.2017SHZDZX01), Shanghai Municipal Science and Technology Major Project (No.2018SHZDZX01) and ZJLab, Shanghai Natural Science Foundation (No. 17ZR1401600) and the National Natural Science Foundation of China (No. 81971583).References
1. Forner A, Llovet JM, Bruix J.
Hepatocellular carcinoma. Lancet (London, England) 2012;379:1245-1255.
2. Radtke A, Nadalin S, Sotiropoulos GC, et al. Computer-assisted
operative planning in adult living donor liver transplantation: a new way to
resolve the dilemma of the middle hepatic vein. World journal of surgery
2007;31:175-185.
3. Avants BB, Epstein CL, Grossman M, Gee JC. Symmetric
diffeomorphic image registration with cross-correlation: evaluating automated
labeling of elderly and neurodegenerative brain. Med Image Anal 2008;12:26-41.
4. Tibshirani R. Regression Shrinkage and Selection via the Lasso.
Journal of the Royal Statistical Society Series B (Methodological)
1996;58:267-288.
5. Chlebus G, Meine H, Moltz J, Schenk A. Neureal Network-Based
Automatic Liver Tumor Segmentation With Random Forest-Based Candidate
Filtering. 2017.
6. Grzegorz Chlebus AS, Jan Hendrik Moltz. Deep learning based
automatic liver tumor segmentation in CT with shape-based post-processing.
2018: Unpublished.