Assessment of U-Net in the segmentation of short tracts: Transferring to clinical MRI routine

Accurately studying structural connectivity requires precise tract segmentation strategies. The U-Net network has been widely recognized for its exceptional capacity in image segmentation tasks and provides remarkable results in large tract segmentation when high-quality diffusion-weighted imaging (DWI) data are used. However, short tracts, which are associated with various neurological diseases, pose specific challenges, particularly when high-quality DWI data acquisition within clinical settings is concerned. Here, we aimed to evaluate the U-Net network ability to segment short tracts by using DWI data acquired in different experimental conditions. To this end, we conducted three types of training experiments involving 350 healthy subjects and 11 white matter tracts, including the anterior, posterior, and hippocampal commissure, fornix, and uncinated fasciculus. In the first experiment, the model was exclusively trained with high-quality data of the Human Connectome Project (HCP) dataset. The second experiment focused on images of healthy subjects acquired from a local hospital dataset, representing a typical clinical routine acquisition. In the third experiment, a hybrid training approach was employed, combining data of the HCP and local hospital datasets. Then, the best model was also tested in unseen DWIs of 10 epilepsy patients of the local hospital and 10 healthy subjects acquired on a scanner from another company. The outcomes of the third experiment demonstrated a notable enhancement in performance when contrasted with the preceding trials. Specifically, the short tracts within the local hospital dataset achieved Dice scores ranging between 0.60 and 0.65. Similar intervals were obtained with HCP data in the first experiment, and a substantial improvement compared to the scores between 0.37 and 0.50 obtained with the local hospital dataset at the same experiment. This improvement persisted when the method was applied to diverse scenarios, including different scanner acquisitions and epilepsy patients. These results indicate that combining datasets from different sources, coupled with resolution standardization strengthens the neural network ability to generalize predictions across a spectrum of datasets. Nevertheless, short tract segmentation performance is intricately linked to the training composition, to validation, and to testing data. Moreover, curved tracts have intricate structural nature, which adds complexities to their segmenting. Although the network training approach tested herein has provided promising results, caution must be taken when extrapolating its application to datasets acquired under distinct experimental conditions, even in the case of higher-quality data or analysis of long or short tracts.