Abdulkhalek Al-Fakih1,2, Abdullah Shazly1,2, Abbas Mohammed1,2, Mohammed Elbushnaq1, Meena Makary1,3, and Mohammed A. Al-masni2
1Department of Biomedical Engineering and Systems, Cairo University, Cairo, Egypt, 2Department of Artificial Intelligence, Sejong University, Seoul, Korea, Republic of, 3MGH/HST Athinoula A. Martinos Center for Biomedical Imaging, Department of Radiology, Massachusetts General Hospital, Harvard Medical School, Charlestown, MA, United States
Synopsis
Keywords: Synthetic MR, Data Analysis, Brain Tumor Segmentation, MR Sequence Synthesis, nnU-net, GANs, Multi Contrast MR, Deep Learning
Motivation: Automated and robust segmentation of brain tumors requires multiple MRI sequences and can expedite neuro-oncological clinical trials.
Goal(s): Our goal was to develop a deep learning model for brain tumor segmentation, even when some MRI sequences are missing.
Approach: We enhanced a GAN with attention modules to synthesize missing sequences and employed an optimized nnU-Net for segmentation using both real and synthesized sequences.
Results: The proposed AI-based model significantly improved brain tumor segmentation, with overall Dice scores increasing from 0.688% when FLAIR is missing to 0.873% using synthesized FLAIR derived from T2, and achieving 0.901% with real FLAIR.
Impact: The developed two-stage deep learning framework,
comprising synthesis and segmentation, enhances segmentation of brain tumors in
MRI, especially when real sequences are unavailable. This advancement
accelerates clinical trials and reduces manual segmentation time, yielding
promising results.
Introduction
Optimal assessment of glioblastoma
necessitates the utilization of information derived from all four MRI
sequences, namely T1, T2, FLAIR, and T1ce. However, acquiring and incorporating
data from all these sequences can be challenging due to the time-consuming and
costly nature of the sequence acquisition process.1 In addition,
artifacts are commonly present in some sequences due to various reasons such as
motion artifacts and magnetic inhomogeneity artifacts. As a result,
accessing complete information from all four sequences may not always be
feasible. This study aims to address the challenge of acquiring complete MRI
sequences, essential for optimal glioblastoma segmentation. We propose a
methodology to synthesize missing MRI sequences, thus ensuring all necessary
data is available for accurate tumor segmentation.Methods
This study proposes a deep learning
pipeline that encompasses identifying and synthesizing missing MRI sequences and
then using these synthesized sequences for brain tumor segmentation, as
illustrated in Figure 1. We utilize the BraTS2021 dataset, which
includes 1251 subjects, to address the problem of missing MRI sequences.
The proposed AI-based approach initiates
by identifying types of brain MRI sequences using an 18-layer residual network.2
Subsequently, we utilize a modified 2D GAN, consisting of a squeeze attention
U-Net generator and a PatchGAN3 discriminator, to synthesize absent MRI
sequences from the existing ones, completing the set, as depicted in Figure 2.
The primary objective of Squeeze-and-Excitation (SE) block4 is to
recalibrate channel responses by explicitly capturing the interdependencies
between them. For generating missing FLAIR sequences, we trained three distinct
GAN models that derived input from T1ce, T1, and T2 images. To obtain a
consolidated loss measure during generator training, the total loss is
calculated as the average of SSIM and MAE, providing insight into both structural
similarity and pixel-wise dependencies.
Next,
the four MRI sequence volumes, including any synthesized missing sequences, are
fed into a self-configuring 3D optimized nnU-Net segmentation model5
for delineating brain tumors (Edema, Enhancing Tumor, and Necrosis). In this
task, a combination of cross-entropy and Dice losses is used as the loss
function, and deep supervision is applied to compute the loss at the last three
decoder layers.Results
In this study, we investigated the
impact of omitting each MRI sequence individually by replacing it with black
images (i.e., zero maps) on the task of brain tumor segmentation. Our findings
revealed that both T1ce and FLAIR images play a crucial role in accurately
segmenting brain tumors compared to the other T1 and T2 sequences. However,
compensating for the contrast agent injected in T1ce poses challenges for deep
learning models. Therefore, our study primarily focuses on synthesizing the
FLAIR sequence. Figure 3 shows some examples of the synthesized FLAIR
sequences generated from T1, T1ce, and T2 sequences.
To demonstrate the effectiveness of
our synthesis method, we compared it to scenarios where sequences are missing
or substituted (i.e., replaced by a black volume or by a copy of T2). Notably,
the absence of the FLAIR sequence significantly affects segmentation compared
to other sequences as shown in Table 1. Visual examples in Figure 4 illustrate the significance of the synthesis process in improving
segmentation accuracy when compared to alternative methods involving black and
copy substitutions.Discussion
Deep learning techniques for brain
tumor segmentation often necessitate the availability of multimodal MRI images,
which might not be doable in clinical practice due to various reasons,
including differences in acquisition protocols and the presence of image
artifacts. This hinders the development and application of such models. This
study demonstrates the effectiveness of GANs in generating missing MRI
sequences, particularly FLAIR images crucial for tumor segmentation. our
proposed GANs using Squeeze Attention U-Net and PatchGAN achieved favorable
results, especially when synthesizing FLAIR from T2 sequences.
The
synthesized FLAIR sequences were integrated with available data and improved
segmentation by 18.5% on average Dice score compared to no FLAIR and by 7%
compared to a copy of the T2 sequence. A multi-model approach, combining
different inputs, outperformed individual models, offering promising solutions
for missing MRI sequences and advancing medical segmentation efforts.Conclusion
Synthesizing missing MRI sequences
can enhance medical segmentation efforts. Promising results were obtained when
the synthesized sequence was combined with the three already-existing sequences
in the brain tumor segmentation model. The absence of the FLAIR sequence
notably impacts segmentation, making its synthesis crucial. Synthesizing the
FLAIR sequence from the T2 sequence has exhibited superior performance in terms
of SSIM and Dice similarity coefficient compared to synthesizing FLAIR from
other sequences.Acknowledgements
This work was supported by the
National Research Foundation of Korea (NRF) funded by the Korean government
(MSIT) (No. RS-2023-00243034).References
[1]
Arti Tiwari, Shilpa Srivastava, and
Millie Pant, “Brain tumor segmentation and classification from magnetic resonance
images: Review of selected methods from 2014 to 2019,” Pattern recognition
letters 131 (2020): 244-260.
[2]
Jean Pablo Vieira de Mello, Thiago
M. Paixão, Rodrigo Berriel, Mauricio Reyes, Claudine Badue, Alberto F. De
Souza, and Thiago Oliveira-Santos, “Deep Learning-based Type Identification of
Volumetric MRI Sequences,” In 2020 25th International Conference on Pattern
Recognition (ICPR), pp. 1-8. IEEE, 2021.
[3]
Phillip Isola, Jun-Yan Zhu, Tinghui
Zhou, and Alexei A. Efros, “Image-to-image translation with conditional
adversarial networks,” In Proceedings of the IEEE conference on computer vision
and pattern recognition, pp. 1125-1134, 2017.
[4]
Jie Hu, Li Shen, and Gang Sun,
“Squeeze-and-excitation networks,” In Proceedings of the IEEE conference on
computer vision and pattern recognition, pp. 7132-7141, 2018.
[5]
Fabian Isensee, Jens Petersen, Andre
Klein, David Zimmerer, Paul F. Jaeger, Simon Kohl, Jakob Wasserthal et al.,
“nnu-net: Self-adapting framework for u-net-based medical image segmentation,”
arXiv preprint arXiv:1809.10486 (2018).