Milletari et al. discuss volumetric, fully convolutional neural networks for medical image segmentation. Concretely, they use the architecture depicted in Figure 1 to segment MRI volumes. Due to the limited training set size, the data was augmented by random deformations and intensity adaptations. The network architecture combines a compression and a decompression part. Both parts consist of several stages, each stage comprising several convolutional layers, followed by adding the input (i.e. to learn the residual) and a downsampling stage (which is implemented by strided convolution in contrast to max pooling). In order to cope with the inbalanced label distribution in the volumes, they propose the dice loss:
where $N$ is the number of voxels, $p_i$ the foreground probability of the prediction and $q_i$ the foreground probability of the ground truth segmentation. As discussed in the paper, the dice loss is differentiable and allows training without assigning weights to the different classes. They provide experimental results on the PROMISE 2012 challenge dataset, see the paper.
Figure 1 (click to enlarge): Illustration of the used network architecture consisting of a compression path (left) and a decompression path (right).
What is your opinion on this article? Let me know your thoughts on Twitter @davidstutz92 or LinkedIn in/davidstutz92.
Milletari et al. discuss volumetric, fully convolutional neural networks for medical image segmentation. Concretely, they use the architecture depicted in Figure 1 to segment MRI volumes. Due to the limited training set size, the data was augmented by random deformations and intensity adaptations. The network architecture combines a compression and a decompression part. Both parts consist of several stages, each stage comprising several convolutional layers, followed by adding the input (i.e. to learn the residual) and a downsampling stage (which is implemented by strided convolution in contrast to max pooling). In order to cope with the inbalanced label distribution in the volumes, they propose the dice loss:
$D = \frac{2\sum_i^N p_i g_i}{\sum_i^N p_i^2 + \sum_i^N g_i^2}$
where $N$ is the number of voxels, $p_i$ the foreground probability of the prediction and $q_i$ the foreground probability of the ground truth segmentation. As discussed in the paper, the dice loss is differentiable and allows training without assigning weights to the different classes. They provide experimental results on the PROMISE 2012 challenge dataset, see the paper.
Figure 1 (click to enlarge): Illustration of the used network architecture consisting of a compression path (left) and a decompression path (right).