1 of 23

Structural Brain MRI Segmentation Beyond FreeSurfer

Ashkan Nejad

Early-Stage Researcher at Royal Dutch Visio and UMCG

welcome, welkom, tere tulemast, tervetuloa, bienvenue, benvenuta, velkommen, добро пожаловать, bienvenidas, 欢迎 , स्वागत हे, 환영, ברוך הבא, hosgeldiniz, أهلا بك, خوش آمدید

2 of 23

Brain MRI Segmentation

  • Brain MRI segmentation is commonly used for:
    • measuring and visualizing the brain's anatomical structures,
    • surgical planning and image-guided interventions,
    • delineating pathological regions,
    • analyzing brain changes [1].

2

2/16/2022

Figure extracted from [2]

3 of 23

  • A software package for the analysis and visualization of structural and functional neuroimaging data [4].

Advantages:

    • Detailed documentation
    • Ubiquitous

Disadvantages:

    • Computational time: 20-24 hours
    • Inability of target set customization

​

3

2/16/2022

[3]

4 of 23

Alternatives

FreeSurfer is not be the only available solution.

    • SLANT,
    • SAU-Net,
    • FastSurfer

Limitation of alternatives: High GPU requirement and difficulty in target set customization

Introducing FLBS framework

4

2/16/2022

5 of 23

SLANT

  • Spatially Localized Atlas Network Tiles (SLANT) was introduced in 2019

Advantages:

    • 120 structures in about 15 minutes
    • Trained networks are available

Shortcomings:

    • Inability of target set customization
    • GPU memory usage

​

5

2/16/2022

[2]

6 of 23

SLANT

6

2/16/2022

7 of 23

SAU-Net

7

2/16/2022

[5]

8 of 23

FastSurfer

8

2/16/2022

[6]

9 of 23

Limitations

  1. Relatively high GPU requirement.
  2. Possible to customize the target set but yet difficult.

​

9

2/16/2022

10 of 23

FLBS (Fast & Light Brain MRI Segmentation)

  • Github.com/arnejad/FLBS

10

2/16/2022

In collaboration with Tehran University of Medical Sciences and Iran’s National Elites Foundation

11 of 23

Evaluation

  • Criterion: Dice Coefficient.

​

  • Gold standard is FreeSurfer results.

​

  • Scenarios:
    1. Train and test

OASIS-3 [8] dataset for both (100 samples for train, 40 for test)

    • Train, transfer and test

OASIS-3 dataset for train and CoRR-BNU1 [9] dataset for test (700 samples train, 50 for test)

11

2/16/2022

Figure extracted from [7]

12 of 23

Comparison – No Transfer Learning

12

2/16/2022

13 of 23

Comparison – Transfer Learning

13

2/16/2022

14 of 23

Comparison – Memory and Time Balance

14

2/16/2022

FLBS

Executed on GPU: Geforce GTX 1080 Ti

CPU: Core-i7 6560U

15 of 23

Discussion

  • Structural brain segmentation is one of the essential steps in diagnosing neurological disorders.
  • FreeSurfer as one of the most used tools has a high time-consumption.
  • Alternative methods reduce the time significantly whit increasing GPU requirement.
  • We compared the performance of the alternative methods and introduced a FLBS to tackle the time-memory balance problem.

​

15

2/16/2022

16 of 23

16

2/16/2022

?

The presentation content will be available on nejad.info

17 of 23

References

[1] Despotović, Ivana, Bart Goossens, and Wilfried Philips. "MRI segmentation of the human brain: challenges, methods, and applications." Computational and mathematical methods in medicine 2015 (2015).

[2] Huo, Yuankai, Zhoubing Xu, Yunxi Xiong, Katherine Aboud, Prasanna Parvathaneni, Shunxing Bao, Camilo Bermudez, Susan M. Resnick, Laurie E. Cutting, and Bennett A. Landman. "3D whole brain segmentation using spatially localized atlas network tiles." NeuroImage 194 (2019): 105-119.

[3] Dale, A.M., Fischl, B., Sereno, M.I., 1999. Cortical surface-based analysis. I. Segmentation and surface reconstruction. Neuroimage 9, 179-194.

[4] https://surfer.nmr.mgh.harvard.edu/fswiki/

[5] Minho Lee et al., “Split-attention u-net: A fully convolutional network for robust multi-label segmentation from brain mri,” Brain Sciences, vol. 10, no. 12, pp. 974, 2020

[6] Leonie Henschel et al., “Fastsurfer-a fast and accurate deep learning-based neuroimaging pipeline,” NeuroImage, vol. 219, 2020.

[7] https://towardsdatascience.com/metrics-to-evaluate-your-semantic-segmentation-model-6bcb99639aa2

[8] LaMontagne, Pamela J., et al. "OASIS-3: longitudinal neuroimaging, clinical, and cognitive dataset for normal aging and Alzheimer disease." MedRxiv (2019).

[9] Zuo, Xi-Nian, et al. "An open science resource for establishing reliability and reproducibility in functional connectomics." Scientific data 1.1 (2014): 1-13.

17

2/16/2022

18 of 23

ML/DL Basic Definitions

Training: Optimizing the ML/DL method based on a part of the data.

​

Inference: Labeling rest of the data using the optimized ML/DL core.

​

Loss function: The main formulated evaluation of the performance to optimize.

​

Dice coefficient: A well-known loss function for segmentation problems.

18

2/16/2022

19 of 23

19

2/16/2022

Courtesy of Devani Cordero, Stanford University

20 of 23

Executing SLANT

20

2/16/2022

# you need to specify the input directory

export input_dir=/home/input_dir

​

# make that directory

sudo mkdir $input_dir

​

# set output directory

export output_dir=$input_dir/output

​

#run the docker

sudo nvidia-docker run -it --rm -v $input_dir:/INPUTS/ -v $output_dir:/OUTPUTS … masidocker/public:deep_brain_seg_v1_1_0 /extra/run_deep_brain_seg.sh

21 of 23

Comparison – Accuracy (Dice Coefficient)

21

2/16/2022

Train and inference on OASIS-3 dataset

Train on OASIS-3 dataset and inference on CoRR-BNU1

22 of 23

Content

  • Structural brain MRI segmentation
  • FreeSurfer
  • Similar Methods
    1. SLANT
    2. SAU-Net
    3. FastSurfer
  • Final Comparison

22

2/16/2022

23 of 23

SAU-Net

  • Split-Attention U-Net (SAU-Net) is a deep-learning architecture that can segment the whole brain.
  • Into 33 structural labels.

​

23

2/16/2022

[5]