Semantic segmentation of cerebrospinal fluid and brain volume with a convolutional neural network in pediatric hydrocephalus-transfer learning from existing algorithms.

backgroundFor the segmentation of medical imaging data, a multitude of precise but very specific algorithms exist. In previous studies, we investigated the possibility of segmenting MRI data to determine cerebrospinal fluid and brain volume using a classical machine learning algorithm. It demonstrated good clinical usability and a very accurate correlation of the volumes to the single area determination in a reproducible axial layer. This study aims to investigate whether these established segmentation algorithms can be transferred to new, more generalizable deep learning algorithms employing an extended transfer learning procedure and whether medically meaningful segmentation is possible.

methodsNinety-five routinely performed true FISP MRI sequences were retrospectively analyzed in 43 patients with pediatric hydrocephalus. Using a freely available and clinically established segmentation algorithm based on a hidden Markov random field model, four classes of segmentation (brain, cerebrospinal fluid (CSF), background, and tissue) were generated. Fifty-nine randomly selected data sets (10,432 slices) were used as a training data set. Images were augmented for contrast, brightness, and random left/right and X/Y translation. A convolutional neural network (CNN) for semantic image segmentation composed of an encoder and corresponding decoder subnetwork was set up. The network was pre-initialized with layers and weights from a pre-trained VGG 16 model. Following the network was trained with the labeled image data set. A validation data set of 18 scans (3289 slices) was used to monitor the performance as the deep CNN trained. The classification results were tested on 18 randomly allocated labeled data sets (3319 slices) and on a T2-weighted BrainWeb data set with known ground truth.

resultsThe segmentation of clinical test data provided reliable results (global accuracy 0.90, Dice coefficient 0.86), while the CNN segmentation of data from the BrainWeb data set showed comparable results (global accuracy 0.89, Dice coefficient 0.84). The segmentation of the BrainWeb data set with the classical FAST algorithm produced consistent findings (global accuracy 0.90, Dice coefficient 0.87). Likewise, the area development of brain and CSF in the long-term clinical course of three patients was presented.

conclusionUsing the presented methods, we showed that conventional segmentation algorithms can be transferred to new advances in deep learning with comparable accuracy, generating a large number of training data sets with relatively little effort. A clinically meaningful segmentation possibility was demonstrated.