Yuval Buchsweiler1,2, Orna Aizenstein3,4, Felix Bokstein3,5,6, Idan Bressler1,2, Netanell Avisdris2,7, Deborah T. Blumenthal3,5, Dror Limon3,8, Dafna Ben Bashat2,3,6, and Moran Artzi2,3,6
1The Iby and Aladar Fleischman Faculty of Engineering, Tel Aviv University, Tel Aviv, Israel, 2Sagol Brain Institute, Tel Aviv Sourasky Medical Center, Tel Aviv, Israel, 3Sackler Faculty of Medicine, Tel Aviv University, Tel Aviv, Israel, 4Division of Radiology, Tel Aviv Sourasky Medical Center, Tel Aviv, Israel, 5Neuro-Oncology Service, Tel Aviv Sourasky Medical Center, Tel Aviv, Israel, 6Sagol School of Neuroscience, Tel Aviv University, Tel Aviv, Israel, 7School of computer science and engineering, Hebrew University of Jerusalem, Israel, Jerusalem, Israel, 8Division of Oncology, Tel Aviv Sourasky Medical Center, Tel Aviv, Israel
Synopsis
Brain
tumor segmentation is highly important for clinical management. We propose
HUMBLe, a hierarchical 3D U-Net for MRI Brain Lesion segmentation architecture.
HUMBLe breaks down the segmentation into its separate classes: enhancing tumor,
edema, and necrotic classes, and uses a classifier to merge the different
segmentation results into a final segmentation mask. Evaluation was performed
on multi-parametric longitudinal local dataset, of patients with Glioblastoma.
Segmentation results obtained by HUMBLe on our cohort improved DICE scores by
7%-16% for the different tumor components, compared to segmentation performed
using 3D U-Net based architecture trained on BraTS2019 and our cohort.
Introduction
High-grade-glioma (HGG) is
an aggressive type of malignant brain-tumor that grows rapidly, usually
requires surgery and radio/chemo-therapy and has poor survival prognosis. Measurements
of changes in tumor size are important for accurate diagnosis and follow-up.
Currently, the most widely used criteria for therapy response assessment in those patients is the
response assessment in neuro-oncology criteria, which relies on 2D manual
measurements, and thus is less accurate and prone to human errors1. Volumetric
measurement is preferable, however in clinical setup, manual volumetric segmentation
of HGG is usually inapplicable as it is highly challenging, time consuming and
user dependent.
In the last few years, the use of deep-learning methods for brain-tumor
segmentation has gained popularity, with a great number of studies conducted on
the BraTS dataset composed of newly diagnosed patients2. Most of
previously suggested methods classify the tumor into its clinically
significant components using multi-class segmentation algorithms. However, applying
those methods for patients' therapy response monitoring is highly challenging
due to the treatment related variability in intensities, presence of surgical
cavities and fuzzy tumor borders that appear in the clinical domain following
tumor resection and chemo/radio/biological therapies.
In this study, we propose HUMBLe, a hierarchical 3D
U-Net based deep-learning segmentation architecture that breaks down the
segmentation into its separate classes and uses a classifier to merge the
different segmentation results into a final segmentation mask.Methods
Input Data- BraTS2019 Dataset: Composed of MRI data of
261 patients with Glioblastoma, annotated with manual segmentation of tumor
area.
- Local
clinical Dataset: Comprised of longitudinal MRI data of six patients with
Glioblastoma (4 scans per patient, total of 24 scans), all following tumor
resection and chemotherapy/radiotherapy. Manual segmentation of tumor area was
done in a similar manner to BraTS dataset.
- Imaging
Data: All cases were composed
of multi-modal MRI, which includes T1 weighted image (WI), T1WI
contrast-enhanced (T1WICE), T2WI and FLAIR modalities. Annotations were comprised
of the Enhancing Tumor, Edema and Necrotic Tumor.
- Data
Preprocessing: Included
image co-registration, bias field correction, background removal, image
cropping (brain region delineation) and
image resizing.
Proposed
method (HUMBLe)A
hierarchical deep-learning segmentation architecture (HUMBLe) was used (Figure
1).
Each
tumor component (Enhancing Tumor, Edema and Necrotic Tumor) was segmented
individually, using a separate network, based on Isensee's et al. 3D U-Net
3. Each network was
trained with complete images with a batch size of 1 and a focal-DICE
4 loss function. For
each network, the ground truth segmentation mask was filtered to include only
the desired class. The segmentation masks of the three separate networks were
fused using a one-vs-rest Linear-SVM
5 classifier to
automatically select the value of each pixel in the final segmentation mask.
Data SplittingBraTS2019
training dataset was split into 80% training and 20% validation datasets. The
second stage of training was performed on our local clinical dataset using
a leave one out method.
Experimental
Scheme-
3D U-Net BraTS Trained:
Training the 3D U-Net
based network over the BraTS2019 dataset, with all 4 modalities as input, while
predicting all class labels. Applying inference on our local clinical dataset.
- 3D U-Net Transfer
Learning: Training the 3D U-Net
based network over the BraTS2019 dataset, with all 4 modalities as input, while
predicting all class labels, followed by fine-tuning the model weights by
training on our dataset.
- HUMBLe: Training the 3 pathways
of the HUMBLe architecture, incorporating the Linear-SVM Classifier (Figure 1)
over the BraTS2019 dataset, followed by fine-tuning the model weights by
training on our dataset.
Evaluation The
network's results were compared using the DICE-Sorensen score for the three
experimental schemes. For the HUMBLe architecture, we determined for each tumor
class separately, which combination of modalities (T1WI, T1WICE, T2WI, FLAIR) results
with the highest DICE scores (on our dataset).
Results
- Segmentation
results for the 3 experimental schemes are given in Table 1. Best segmentation
results were achieved by HUMBLe using all 4 modalities as input for the Enhancing and Necrotic Tumor classes, and only using the T2WI and FLAIR for the Edema class (Figure
1). The
fused segmentation map of the model showed significant improvement in DICE
scores with 7% for the Whole Tumor (Enhancing+Edema+Necrotic), 16% for the Tumor Core (Enhancing+Necrotic) and 15% for the Enhancing Tumor, in our cohort over
the other two models (Table 1).
- Figure
2 demonstrates segmentation comparison for a patient with GBM. As can be
seen, the best results were obtained by HUMBLe, which copes better with the
challenge of tumor segmentation given the tumor intensity variabilities and
fuzzy borders following therapies.
- Figure
3 shows an example of longitudinal segmentation result for a patient treated
with Bevacizumab (between time points 2 and 3), which in many cases results in challenges
to radiological therapy response assessment in those patients, due to the
substantial change in tissue intensities6. The success of HUMBLe
in tumor segmentation in those cases demonstrates the clinical potential of
this tool.
Discussion
In
this work we proposed the HUMBLe architecture for volumetric assessment of
patients with Glioblastoma. HUMBLe achieved the best segmentation results in
terms of DICE score for patients following tumor resection and
chemo/radio/biological therapies and thus, it may be integrated in clinical
practice for longitudinal therapy response assessment for patients with
Glioblastoma.Acknowledgements
No acknowledgement found.References
- Wen,
P.Y., et al., Updated response assessment
criteria for high-grade gliomas: response assessment in neuro-oncology working
group. Journal of clinical oncology, 2010. 28(11): p. 1963-1972.
-
Bakas,
S., et al., Identifying the best machine
learning algorithms for brain tumor segmentation, progression assessment, and
overall survival prediction in the BRATS challenge. arXiv preprint
arXiv:1811.02629, 2018.
-
Isensee,
F., et al. Brain tumor segmentation and
radiomics survival prediction: Contribution to the brats 2017 challenge. in
International MICCAI Brainlesion Workshop.
2017. Springer.
- Abraham,
N. and N.M. Khan. A novel focal tversky
loss function with improved attention u-net for lesion segmentation. in 2019 IEEE 16th International Symposium on
Biomedical Imaging (ISBI 2019). 2019. IEEE.
-
Cortes,
C. and V. Vapnik, Support-vector
networks. Machine learning, 1995. 20(3):
p. 273-297.
-
Arevalo,
O.D., et al., Assessment of glioblastoma response
in the era of bevacizumab: longstanding and emergent challenges in the imaging
evaluation of pseudoresponse. Frontiers in neurology, 2019. 10: p. 460.