Anna Petrova1,2, Assunta Dal-Bianco2,3, Eva Niess1, Nik Krajnc2,3, Wolfgang Bogner1, Günther Grabner4, Paul Rommer2,3, and Stanislav Motyka1
1High Filed MR Center, Department of Biomedical Imaging and Image-guided Therapy, Medical University of Vienna, Vienna, Austria, 2Department of Neurology, Medical University of Vienna, Vienna, Austria, 3Comprehensive Center for Clinical Neurosciences & Mental Health, Medical University of Vienna, Vienna, Austria, 4Department of Medical Engineering, Carinthia University of Applied Sciences, Klagenfurt, Austria
Synopsis
Keywords: Diagnosis/Prediction, Machine Learning/Artificial Intelligence
Motivation: The effective treatment of Multiple sclerosis (MS) requires reliable estimates of lesion load and hence precise lesion detection over time. However, current lesion load estimation is either qualitative or too time-consuming.
Goal(s): Our study automates MS lesion segmentation by training DeepMedic for application to 7T multi-contrast MRI data of MS patients.
Approach: Training with all four contrasts achieved the best results compared to Lesion Segmentation Tool (LST)—a conventional/non-deep-learning SPM-based MS lesions segmentation approach.
Results: Our study highlights potential for automating MS lesion detection/segmentation for 7T multi-contrast MRI data, underscoring the importance of accurate ground truth data and high-quality databases for improved detection accuracy.
Impact: The
results of this research will impact the user-independent
detection/segmentation of multiple sclerosis lesions, making manual
assessment by clinicians obsolete and enable fully automated
monitoring of lesions load as a quantitative radiological marker of
disease progression.
Introduction
Multiple
sclerosis (MS) requires early, precise lesion detection for
effective treatment planning and prognosis. Manual segmentation is
challenging due to small lesion sizes and time-consuming. Recent
advancements in deep learning hold potential for automating lesion
segmentation in various tasks[1,2]. Nonetheless, MS presents a continued
challenge for segmentation accuracy, especially in clinical MRI
(FLAIR/MPRAGE at 1.5T and 3T). Our study introduces automated lesion
segmentation using 7T multi-contrast MRI data with the DeepMedic
network, employing a combination of Binary Cross-Entropy (BCE) and
Volume-level sensitivity–specificity (VSS) loss function[3,4].Methods
Experimental
Data
The
database contained multi-contrast 7 Tesla MRI data of 31 MS patients
(age 42±15; 26 RRMS and 5 SPMS; EDSS 1,55±3.25). Four different MRI
contrasts were measured for each subject: FLAIR, MP2RAGE, QSM, and
SWI. Parameters of sequences and some pre-processing steps presented
in Table 2.
Training
and
testing
data
For
training of the DeepMedic network, 23 subjects from the database were
used, along with three subjects for validation. Five subjects were
reserved for testing the performance of the neural network. Reference
masks of lesions were manually segmented by an experienced
neurologist (considered gold standard).
Neural
network architecture
The
network consists of two main components: a 3D convolutional neural
network (CNN) for feature extraction and a fully connected
conditional random field (CRF). To enhance the effectiveness of MS
lesion detection, it was proposed to combine two loss functions, BCE
and VSS [4]. This combination of loss functions allows to assess the
performance of MS lesion detection not only at the voxel level but
also assesses sensitivity and specificity of lesion detection at the
volume level.
Network
training
The
network was trained using various combinations of contrasts.
Initially, only one contrast, either FLAIR or MP2RAGE, was used.
Subsequently, training was performed for two different combinations
of contrasts: FLAIR+MP2RAGE and all four contrasts. Each training
session involved 50 epochs, taking approximately 72 hours to
complete. Training and network inference were conducted on an Nvidia
DGX station using a single graphics card (Nvidia Tesla V100 32 GB).
The network configurations used in our work were recommended by the
authors of DeepMedic [3].
Evaluation
The
prediction of lesion mask by different versions of network were
evaluated alongside the segmentation with a conventional established
approach—Lesion Segmentation Tool (LST) [5]—an extension of
SPM12. LSM segmentation used only FLAIR images and did not involve
MP2RAGE and others contrasts.
To
measure sensitivity and accuracy, assessments are made at the lesion
level based on confusion matrix. To account for both location and
size of segmentation, the Dice Similarity Coefficient (DSC) was used.
Additionally, Intersection Over Union and Matthew's correlation
coefficient were calculated.Results
The
assessment results after network training on various contrast
combinations are presented in Table 1, along with the results from
LST (results presented for one patient).
Results after training using a single contrast show that FLAIR
achieves slightly better lesion detection accuracy (DSC=0.720) than
MP2RAGE (DSC=0.705). An example for false positives (FP) lesion
segmentation after training with these two contrasts are shown in
Figure 2. However, the network achieves better sensitivity and
accuracy when trained with all four contrasts (DSC=0.738), as
indicated in Table 1.
In
addition, a comparison was conducted between DeepMedic and SPM12 LST
(Figure 1). The data obtained from this comparison indicates that
accuracy of DeepMedic trained with FLAIR contrast (DSCmean=0.707)
is higher than for SPM12 LST (DSCmean=0.659).
A detailed comparison is provided in Table 2 (result
presented for three patients).
Deep
learning lesion segmentation was not only more accurate, but also
significantly faster (~3min per patient) than the manual segmentation
performed by a clinician (4-5 hours per case).Discussion and Conclusion
In
this study, we performed automated MS lesion segmentation using 7T
multi-contrast MRI data with the DeepMedic network. We explored
various training combinations, including single-contrast (FLAIR or
MP2RAGE), dual-contrast (FLAIR and MP2RAGE), and all four contrasts.
These combinations provide confidence in distinguishing true positive
lesions, though expert assessment is still necessary. Our gold
standard relied on a single experienced neurologist's segmentation,
that make this limitation. Despite this, four-contrast 7T
multi-contrast MRI data-based segmentation with DeepMedic produced
excellent results.
Additionally,
we demonstrated the segmentation advantages of this neural network
over SPM12 LST. Nevertheless, challenges in the gold standard
contribute to less-than-perfect results, and an expanded dataset with
more patients and expert manual segmentations is necessary. Employing
two loss functions and high-quality databases can significantly
enhance the accuracy of MS lesion detection and segmentation for
clinical use.Acknowledgements
No acknowledgement found.References
- Gros
C., De Leener B, et al. Automatic
segmentation of the spinal cord and intramedullary multiple
sclerosis lesions with convolutional neural networks. Neuroimage.
2019 Jan 1;184:901-915. doi: 10.1016/j.neuroimage.2018.09.081. Epub
2018 Oct 6. PMID: 30300751; PMCID: PMC6759925.
-
Zeng
C, Gu L, Liu Z, Zhao S. Review of Deep Learning Approaches for the
Segmentation of Multiple Sclerosis Lesions on Brain MRI. Front
Neuroinform. 2020 Nov 20;14:610967. doi: 10.3389/fninf.2020.610967.
PMID: 33328949; PMCID: PMC7714963.
-
K.
Kamnitsas, C. Ledig, et.al. Efficient
multi-scale 3D CNN with fully connected CRF for accurate brain
lesion segmentation. Medical Image Analysis,Volume 36, 2017, Pages
61-78, ISSN 1361-8415, https://doi.org/10.1016/j.media.2016.10.004.
- Huang
Y, Bert C, et all. Deep
learning for brain metastasis detection and segmentation in
longitudinal MRI data. Med Phys. 2022 Sep;49(9):5773-5786. doi:
10.1002/mp.15863. Epub 2022 Jul 22. PMID: 35833351.
- Egger
et al. (2017). MRI FLAIR lesion segmentation in multiple sclerosis:
Does automated segmentation hold up with manual annotation?
NeuroImage. Clinical, 13, 264–270.
http://doi.org/10.1016/j.nicl.2016.11.020