Multiple sclerosis (MS) segmentation is a crucial task that helps to monitor the progression of that condition and to investigate how efficient is the treatment provided to a patient. Convolutional Neural Networks (CNN) have been successfully employed in MS lesion segmentation in recent years, but still have problems in segmenting voxels in the boundaries of the lesions. In this work, we present a modified CNN that assign more importance in learning hard-to-classify voxels close to the boundaries of the MS lesions. During the training process, we performed a stratified sampling to dynamically increase the penalization of voxels in the neighborhood around MS lesions boundaries. We prove that the stratified sampling strategy increases the representation of voxels near to the neighborhood of the edges and retrieves more accurate results in terms of Dice similarity coefficient compared to existing methods that uses uniform sampling. To test our approach, the 2015 Longitudinal MS Lesion Segmentation Challenge dataset was used, obtaining Dice > 0.7, which is comparable to the performance of human experts.
- Convolutional neural networks
- focal loss
- lesions segmentation
- magnetic resonance imaging
- multiple sclerosis