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

2185 

Abstract Submission 

Hohana Konell¹, Antonio dos Santos², Carlos Ernesto Garrido Salmon¹

¹InBrain Lab, Department of Physics, Faculty of Philosophy, Sciences and Letters of Ribeirão Preto, Ribeirão Preto, São Paulo, ²Department of Medical Imaging, Hematology and Clinical Oncology, Ribeirão Preto Medical School, Ribeirão Preto, São Paulo

Hohana Konell, M.S.  
InBrain Lab, Department of Physics, Faculty of Philosophy, Sciences and Letters of Ribeirão Preto
Ribeirão Preto, São Paulo

Antonio dos Santos  
Department of Medical Imaging, Hematology and Clinical Oncology, Ribeirão Preto Medical School
Ribeirão Preto, São Paulo

Carlos Ernesto Garrido Salmon, PhD  
InBrain Lab, Department of Physics, Faculty of Philosophy, Sciences and Letters of Ribeirão Preto
Ribeirão Preto, São Paulo

Accurately studying structural connectivity requires precise tract segmentation strategies (Zhang, 2021). The U-Net network has been widely recognized for its exceptional capacity in image segmentation tasks (Ronnenberg, 2015). It has demonstrated remarkable results in segmenting large tracts using high-quality diffusion-weighted imaging (DWI) data (Wasserthal, 2018). However, short tracts, which are associated with various neurological diseases, pose specific challenges, particularly when considering the DWI data acquisition within clinical settings. The objective of this work was to evaluate the capability of the U-Net network in segmenting short tracts using DWI data acquired in different experimental conditions.

To accomplish this, we conducted three different types of training experiments with a total of 350 healthy subjects and 11 white matter tracts, including anterior, posterior, and hippocampal commissure, fornix, and uncinate fasciculus. In the first experiment, the model was exclusively trained using high-quality data from the Human Connectome Project (HCP) dataset, presenting 270 gradient directions and 3 bvalues (1000, 2000, 3000 s/mm²). The second experiment focused on images of healthy subjects acquired from a local hospital dataset (CAAE - 08219712.7.0000.5407), representing a typical clinical routine acquisition with 32 gradient directions and b = 1000 s/mm². In the third experiment, a hybrid training approach was employed, combining images from the HCP and local hospital datasets.

The utilized architecture for this study was the 2D U-Net, originally proposed by Wasserthal et al. (2018). In this case, the input consisted of a 2D image with dimensions of 144x144 voxels and 9 channels representing the orientation information of the fiber. To achieve this, the images were sliced into three different orientations: axial, coronal, and sagittal. Consequently, three separate networks were trained, with one dedicated to each orientation. Reference labels were created using pre-defined regions of interest based on existing literature (Pinto, 2020). The individual tracts were transformed into binary masks to establish the ground truth.

To evaluate the performance of the trained model in each experiment, we conducted tests on two separate sets of subjects. The first test was performed on 60 unseen subjects from the HCP dataset and the second on 60 unseen subjects from the local hospital dataset, ensuring that the model was evaluated on data that it had not been trained on. Dice score was used to evaluate the accuracy prediction.

The third experiment showcased significant visual improvement compared to prior trials (Figure 1). Training exclusively on the public dataset posed challenges for tract reconstruction in the local hospital dataset, but the third experiment delivered the most promising results. Particularly, short tracts within the local hospital data achieved dice scores ranging from 0.60 to 0.75 (Figure 2). Additionally, substantial enhancement was observed for the HCP dataset compared to training solely with high-quality data.

This outcome strongly indicates that the fusion of datasets from various sources, coupled with resolution standardization, significantly fortifies the neural network's capacity to generalize predictions across a spectrum of datasets. It's crucial, however, to recognize that the performance of short tract segmentation is intricately linked to the composition of the training, validation, and testing data. Moreover, the segmentation of shorter and intricately curved tracts introduces added complexities due to their intricate structural nature. Although this approach has shown promising results, caution is essential when extrapolating its application to datasets acquired under distinct experimental conditions, even when dealing with higher-quality data or analyzing long or short tracts.

Classification and Predictive Modeling ²

Segmentation and Parcellation

White Matter Anatomy, Fiber Pathways and Connectivity ¹

Computational Neuroscience

Data analysis

Machine Learning

Modeling

MRI

Segmentation

Tractography

White Matter

WHITE MATTER IMAGING - DTI, HARDI, DSI, ETC