Deep learning ensemble for brain tissue segmentation

A 2D and 3D CNN Ensemble Approach for MRI Segmentation


Authors: Frederik Hartmann, Xavier Beltran Urbano
Code: Github
Report: Github

*This project was carried out within the scope of the Medical Image Segmentation course taught by Prof. Dr. Xavier Lladó.

December 2023



Dataset

The dataset used in this study is the IBSR18, containing 18 T1-weighted scans of normal subjects from the Internet Brain Segmentation Repository (IBSR). It includes preprocessed scans with a 1.5 mm slice thickness and ground truth segmentation for white matter (WM), gray matter (GM), and cerebrospinal fluid (CSF).

Example of the IBSR18 Dataset

Methodology

Our approach can be divided into 2 sections:

· Preprocessing

First, normalization was implemented using a robust z-normalization technique, chosen due to the non-Gaussian distribution and the presence of outliers in some data. This involves adjusting the intensity values by subtracting the mean (calculated using the 25th and 75th quantiles) and dividing by the standard deviation, calculated over the same quantile range. Additionally, data augmentation was performed through random flips and rotations of the original images to enhance algorithm reliability. Finally, selective slice selection was employed for multi-class segmentation, prioritizing slices containing cerebrospinal fluid (CSF) to address class imbalance.

· Training

For the training of this approach, several architectures, such as U-Net, Res-U-Net, Dense-U-Net, and SegResNet, have been used. We have also investigated the effect of utilizing distinct image planes (axial and coronal). Additionally, both 2D and 3D approaches have been analyzed. In Table 1, you can observe all the networks that we have trained independently. After the single trainings, we have ensembled the networks in different configurations (see Table 2).

Results

Both quantitative and qualitative results are presented in this section. They are as follows:

· Quantitative Results

In the following tables, we can observe the results obtained for the single trainings and the different ensemble methods carried out in this project. The metrics utilized to evaluate the segmentations are the Dice Coefficient (DSC) and the Hausdorff Distance (HD).

Single Model Results

Table 1: Single model results on the validation set.
Model CSF Dice GM Dice WM Dice Mean Dice CSF HD GM HD WM HD Mean HD
2D Coronal U-Net 0.878 0.937 0.933 0.917 39.352 11.344 10.443 20.380
2D Coronal Dense U-Net 0.899 0.937 0.938 0.925 17.168 12.199 8.149 12.502
2D Coronal Multi-U-Net 0.890 0.935 0.936 0.920 26.234 13.391 8.422 16.016
2D Coronal Res-U-Net 0.882 0.931 0.931 0.915 21.894 12.000 10.905 14.933
2D Axial U-Net 0.868 0.929 0.922 0.906 26.598 9.876 9.887 15.454
2D Axial Dense-U-Net 0.868 0.920 0.920 0.902 27.137 11.281 10.580 16.333
2D Axial Multi-U-Net 0.876 0.923 0.926 0.908 30.938 10.546 9.872 17.119
2D Axial Res-U-Net 0.866 0.925 0.921 0.904 23.733 21.277 10.113 18.375
2D Seg-Res-Net 0.877 0.933 0.935 0.915 13.540 9.977 9.449 10.989
3D U-Net 0.882 0.942 0.942 0.922 16.202 12.864 11.574 13.486
3D Seg-Res-Net 0.888 0.935 0.937 0.921 15.198 10.367 9.541 11.702
SynthSeg 0.812 0.829 0.888 0.843 29.822 8.353 12.066 16.747

Ensemble Results

Table 2: Ensemble results on the validation set.
Model CSF Dice GM Dice WM Dice Mean Dice CSF HD GM HD WM HD Mean HD
The Coronal Ensemble Mean 0.895 0.939 0.939 0.925 18.508 9.630 7.783 11.974
The Coronal Ensemble Maximum 0.893 0.939 0.939 0.923 23.860 9.811 8.843 14.171
The Coronal Ensemble Majority 0.890 0.939 0.937 0.922 19.123 11.465 7.564 12.717
The Axial Ensemble Mean 0.884 0.930 0.928 0.914 17.121 10.704 9.127 12.317
The Axial Ensemble Maximum 0.881 0.930 0.927 0.913 23.055 10.655 9.782 14.498
The Axial Ensemble Majority 0.877 0.930 0.925 0.911 22.114 10.946 9.277 14.112
The Coronal + Axial Mean 0.897 0.939 0.938 0.925 16.410 8.902 9.095 11.469
The Coronal + Axial Maximum 0.893 0.938 0.937 0.923 21.901 10.270 9.370 13.847
The Coronal + Axial Majority 0.894 0.940 0.938 0.924 16.611 9.811 8.653 11.692
The Multidimensional Ensemble Mean 0.904 0.945 0.948 0.932 11.918 8.730 7.660 9.436

· Qualitative Results

Here, an example of the segmentation obtained from the ensemble with the best performance (the Multidimensional Ensemble Mean) is presented in Fig 2.

Comparison of segmentation results of the best-performing ensemble: The Multidimensional Ensemble. Displayed are axial slices (left), sagittal slices (middle), and coronal slices (right)

Conclusion

The study clearly illustrates the efficacy of an ensemble methodology that synergizes 2D and 3D convolutional neural networks (CNNs) for segmenting brain tissue. This innovative approach benefits significantly from leveraging various orientations of 2D slices in combination with both 2D and 3D models. Among the various techniques explored, the ensemble method, especially the mean of probabilities technique, stands out for its exceptional robustness and precision in results. Future scope would include working on dealing with the imbalance problem since, even after obtaining excellent performance for all tissues, it can be appreciated that the results obtained for CSF are slightly lower than the WM and GM ones.