Table 1.
Summary of MRI sequence parameters for manual segmentation dataset (Dataset A).
Table 2.
Summary of MRI sequences parameters for lesion count dataset (Dataset B).
Fig 1.
The model has five depths of encoder and decoder. Blue arrows represent 2D convolution layers with batch normalization and rectified linear unit, red arrows represent max pooling layers, and green arrows represent up sampling layers. Brown arrows are skip connections between the encoder and decoder networks. MRI contrasts and probability maps are the input, and the segmented lesion map is the output.
Table 3.
Model performance for test cases of Dataset A with different loss function.
Fig 2.
Predicted images for different loss functions.
Each row represents image slices from different MRI sequences. The first column represents T1 post-contrast images and the second column represents the enlarged T1 post-contrast images around the lesion of interest. The other columns are performances of UNet and random forest model for different loss functions. Red regions are false negative segmentation, green areas are true positive, and blue regions are false positive regions. The white arrows are used to identify the lesion of interest. The first two rows (a and b) show example of lesions which are identified by all loss functions. The third row (c) shows change of lesion segmentation volume by loss functions. Fourth and fifth rows (d and e) show the improvement of segmentation due to bootstrapping loss function. The last row (f) shows example of false detection.
Fig 3.
Predicted images for different input sequences.
Each row represents image slices from different MRI sequences. The first column represents T1 post-contrast images and the second column represents the enlarged T1 post-contrast images around the lesion of interest. The other columns are performances of UNet and random forest model for different input sequences, which are tabulated below. The red regions are false negative segmentation, green is an overlap, and blue regions are false positives. The white arrow shows the lesion of interest. The first row (a) shows an example of lesion detection using any input while the second row (b) shows change in lesion segmentation volume due to different input. The third row (c) shows a lesion that cannot be detected using only T1 post-contrast. The fourth row (d) shows an example of lesion detection where T1 pre-contrast, T1 post-contrast and FLAIR are most influential to detect. The fifth row (e) shows a lesion that cannot be detected without T2.
Table 4.
Effect of different input combinations of MRI sequences from test cases of Dataset A.
Table 5.
Confusion matrix lesion count results on Dataset B.