Tractography, which refers to computational representation of the white matter pathways obtained through diffusion magnetic resonance imaging (dMRI) [1], requires precise tract segmentation strategies when brain structural connectivity is studied. White matter segmentation allows both healthy and abnormal brain microstructures to be identified and characterized in vivo, providing valuable insights for diagnosis, treatment, and surgical planning [[2], [3], [4]]. However, for this type of analysis to be reliable, tracts must be accurately delineated, which demands time and anatomical and imaging processing knowledge. This blights the prospects of reproducibility because delineation methodologies may vary across researchers. To address these constraints and to enhance segmentation, automatic segmentation strategies have been investigated [5,6].

Although virtual dissection combining inclusion and exclusion regions of interest (ROIs) remains the gold standard, automatic segmentation strategies are being explored to reduce analysis time and variability across operators [[7], [8], [9]]. These strategies can be categorized into three types: ROI-based segmentation, clustering-based segmentation, and direct segmentation. ROI-based segmentation approaches use anatomical priors to define ROIs that guide fiber segmentation [10]. Clustering-based strategies group fibers into coherent clusters by assigning labels to them [11]. Direct strategies aim to obtain segmentation from volumetric data, without involving intermediate processes [12,13]. While ROI-based segmentation strategies are employed the most often, the advent of artificial intelligence has opened new possibilities for direct segmentation, offering improved reproducibility.

Deep learning methods have already demonstrated significant application in dMRI and great potential in tractography segmentation [8,14,15]. For automated bundle segmentation involving deep learning, these methods can be divided into two main groups based on the input data of the network, voxel-based segmentation and streamline-based segmentation. In voxel-based approaches, segmentation is predicted by using orientation information derived from fiber tracts [16], whereas streamline-based segmentation strategies employ pre-defined fiber features to achieve bundle segmentation [14].

The U-Net network stands out as the most applied and renowned architecture in convolutional neural networks for segmentation tasks [17]. The unique ability of this network to process entire imaging volumes and to produce segmentation maps at the output has contributed to its popularity. When the U-Net network is used to segment specific tracts, it provides exceptional results, especially when large tracts and high-quality diffusion-weighted imaging (DWI) data are concerned [[18], [19], [20]].

In the study conducted by Wasserthal et al. [18], the authors employed stacked 2D U-Net models in a supervised manner to segment tracts automatically, considered the state of the art in direct white matter segmentation. Although this strategy excelled in segmenting larger tracts, it performed less satisfactorily for short tracts. In addition, the training dataset was sourced from high-quality data, which prevented the models from being generalized to lower-quality clinical data, thereby impacting their predictive accuracy.

Automatic segmentation strategies have succeeded particularly in the case of large tracts, such as the corticospinal tract and corpus callosum [6,21]. However, short tracts, including the fornix and anterior commissure, have been a challenge for neural networks and have been less explored. These short tracts are associated with various neurological diseases; for instance, Alzheimer's disease affects the uncinate fasciculus, hippocampus, and anterior commissure [[22], [23], [24]], while epilepsy is linked to abnormalities in the hippocampus, fornix, and uncinate fasciculus [[25], [26], [27]]. Moreover, the anterior and posterior commissures serve as essential landmarks in neurosurgery [28,29], which further underscores the need to automate segmentation of these short tracts for clinical applications.

Even though deep learning methods perform impressively, several factors complicate their generalizability, especially when it comes to image acquisition in hospitals. Most studies in the field of white matter tract segmentation have primarily focused on the Human Connectome Project (HCP) dataset. When these methods are applied to clinical acquisitions, they often use datasets obtained from patients with specific neurological diseases or downgrade high-quality data to match the clinical acquisition settings [8,18,20].

The images obtained during routine hospital procedures tend to have different quality from the high-quality images that constitute most public datasets. This discrepancy poses a significant challenge mainly due to time constraints imposed by clinical workflows. Variability in image acquisition protocols, preprocessing pipelines, and brain anatomy adds further complexity given that these factors can influence the resulting tractography [30]. Consequently, addressing these challenges is key to ensuring that deep learning methods are effective and reliable in a clinical setting.

Here, we aimed to evaluate the U-Net network ability to segment short tracts by using DWI data acquired in different experimental conditions. We studied five short tracts—the anterior, posterior, and hippocampal commissure, fornix, and uncinate fasciculus. We obtained the reference segmentation by employing a semi-automatic ROI-based segmentation strategy and applied it to 175 HCP subjects and 175 subjects at a local hospital. This strategy enabled us to train and to evaluate the U-Net network by using conventional clinical images.