Deep Learning Approach for Automated White Matter Fiber Bundle Segmentation
Poster No:
2202
Submission Type:
Abstract Submission
Authors:
Sebastián Navarrete Caro1, Natalia Vidal1, Cecilia Hernández1, Josselin Houenou2, Cyril Poupon2, Jean-François Mangin2, Pamela Guevara1
Institutions:
1Universidad de Concepción, Concepcion, Región del Bio-Bio, 2Université Paris-Saclay, CEA, CNRS, Neurospin, Gif-sur-Yvette, France
First Author:
Co-Author(s):
Introduction:
The segmentation of white matter (WM) tracts based on Diffusion magnetic resonance imaging, known as virtual dissection, is crucial for studying brain structural connectivity and analyzing neurodegenerative diseases (d'Albis et al., 2018), as well as planning brain surgeries to minimize harm (Essayed et al., 2017), among other applications.
Manual segmentation, the gold standard, involves the selection regions of interest and fibers by an expert. This method is time-consuming and presents variability among experts (Rheault et al., 2022).
Automatic WM bundle segmentation methods use anatomical information such as an anatomical atlas (Wassermann et al., 2016) or a WM bundle atlas (Guevara et al. 2012b) and, in general, can label massive tractography datasets quickly with reproducibly results. The method (Vázquez et al. 2019, Guevara et al., 2012b), uses centroids from a WM bundle atlas as reference fibers to identify those fibers from a new subject with a shape and location similar to the atlas fibers. This method has proven its efficacy in several clinical studies [Buyukturkoglu et al., 2022]. It uses a pairwise fiber distance, which is sensitive to differences in fiber shape and length that can lead to the loss of bundle fibers.
In recent years, emerging deep learning methods based on autoencoder like FINTA (Legarreta et al., 2021), TractoFormer (Zhang et al., 2022) and GESTA (Legarreta et al., 2023) aim to enhance tractography data processing. But still, there is a research gap on applying these find of models for automatic WM bundle segmentation. We propose a new method based on a WM bundle atlas to process tractography data in the latent space of an autoencoder trained with a HARDI database.
Manual segmentation, the gold standard, involves the selection regions of interest and fibers by an expert. This method is time-consuming and presents variability among experts (Rheault et al., 2022).
Automatic WM bundle segmentation methods use anatomical information such as an anatomical atlas (Wassermann et al., 2016) or a WM bundle atlas (Guevara et al. 2012b) and, in general, can label massive tractography datasets quickly with reproducibly results. The method (Vázquez et al. 2019, Guevara et al., 2012b), uses centroids from a WM bundle atlas as reference fibers to identify those fibers from a new subject with a shape and location similar to the atlas fibers. This method has proven its efficacy in several clinical studies [Buyukturkoglu et al., 2022]. It uses a pairwise fiber distance, which is sensitive to differences in fiber shape and length that can lead to the loss of bundle fibers.
In recent years, emerging deep learning methods based on autoencoder like FINTA (Legarreta et al., 2021), TractoFormer (Zhang et al., 2022) and GESTA (Legarreta et al., 2023) aim to enhance tractography data processing. But still, there is a research gap on applying these find of models for automatic WM bundle segmentation. We propose a new method based on a WM bundle atlas to process tractography data in the latent space of an autoencoder trained with a HARDI database.
Methods:
Our method employs a convolutional autoencoder to project data into the latent space. We adopt the FINTA autoencoder structure (Legarreta et al., 2021) with minor modifications: replacing ReLU layers with Leaky ReLU layers for processing Talairach space data and increasing the latent space size from 32 to 128 values for better fiber description.
The autoencoder was trained with 5 tractography datasets from the HARDI ARCHI database (Poupon et al., 2012), in Talairach space (∼ 1 million fibers per subject). We used the Adam optimizer with a mean squared error loss. Hyperparameters were tuned using Bayesian search, with fixed values of learning rate of 2.88e-04 and a weight decay of 4.61e-05.
For segmentation, we utilize a deep WM (DWM) bundle atlas (Guevara et al., 2012b) in Talairach space, composed of 36 fiber fascicles, transformed to the latent space. A radius neighbors classifier was trained with different search distances for each fascicle to identify the atlas bundles from new tractography datasets by comparing radial distance with reference data. The proposed segmentation method was applied to a clinical database of 37 male subjects, with 19 high-functioning Autism Spectrum Disorder (ASD) patients, and 18 controls (d'Albis et al., 2018). Mean WM microstructural measures (ADC, FA, and GFA) for each bundle served as input for an SVM classification algorithm (Fig. 1).
The autoencoder was trained with 5 tractography datasets from the HARDI ARCHI database (Poupon et al., 2012), in Talairach space (∼ 1 million fibers per subject). We used the Adam optimizer with a mean squared error loss. Hyperparameters were tuned using Bayesian search, with fixed values of learning rate of 2.88e-04 and a weight decay of 4.61e-05.
For segmentation, we utilize a deep WM (DWM) bundle atlas (Guevara et al., 2012b) in Talairach space, composed of 36 fiber fascicles, transformed to the latent space. A radius neighbors classifier was trained with different search distances for each fascicle to identify the atlas bundles from new tractography datasets by comparing radial distance with reference data. The proposed segmentation method was applied to a clinical database of 37 male subjects, with 19 high-functioning Autism Spectrum Disorder (ASD) patients, and 18 controls (d'Albis et al., 2018). Mean WM microstructural measures (ADC, FA, and GFA) for each bundle served as input for an SVM classification algorithm (Fig. 1).
·Fig. 1. Scheme of the proposed method pipeline.
Results:
To evaluate the proposed segmentation method (S2), the results were compared with the segmentation method (Vázquez et al., 2019) (S1). This comparison was conducted using the 37 subjects from the ASD database. The proposed segmentation algorithm (S2), on average, achieved a better recovery for most of the 36 fascicles compared to S1, and visually reveals a clear improvement in the coverage of segmentations, especially for the thalamic radiations as shown in Fig. 2a. Results from the classification of ASD show an improvement of the accuracy from 73% to 87%.
·Fig. 2 a) Segmented fiber for each fascicle of the left hemisphere. b) Visualization of the left hemisphere fascicle. c) Results for ASD classification using SVM over the WM microstructural features.
Conclusions:
The proposed algorithm improves the quality of WM bundle segmentation in terms of the number of fibers and coverage. Also, a positive impact of the methods was observed for the classification of patients with ASD. Future work will improve the method parameter setting and validation.
Modeling and Analysis Methods:
Classification and Predictive Modeling
Segmentation and Parcellation 2
Neuroanatomy, Physiology, Metabolism and Neurotransmission:
White Matter Anatomy, Fiber Pathways and Connectivity 1
Keywords:
Autism
Machine Learning
MRI
Tractography
White Matter
WHITE MATTER IMAGING - DTI, HARDI, DSI, ETC
1|2Indicates the priority used for review
Provide references using author date format
Essayed W I (2017) White matter tractography for neurosurgical planning: A topography-based review of the current state of the art, NeuroImage: Clinical, 15, pp. 659–672.
Guevara P (2012) Automatic fiber bundle segmentation in massive tractography datasets using a multi-subject bundle atlas, NeuroImage, 61(4), pp. 1083–1099.
Legarreta J H (2023) Generative Sampling in Bundle Tractography using Autoencoders (GESTA), Medical Image Analysis, 85, p. 102761.
Legarreta J (2021) Filtering in tractography using autoencoders (FINTA), Medical Image Analysis, 72, p. 102126.
Poupon C. (2012), Connect/archi: an open database to infer atlases of the human brain connectivity’, ESMRMB 272, 2012.
Rheault F (2022) Tractostorm 2: Optimizing tractography dissection reproducibility with segmentation protocol dissemination, Human Brain Mapping, 43(7), pp. 2134–2147.
Vázquez A (2019) ‘Parallel Optimization of Fiber Bundle Segmentation for Massive Tractography Datasets’, in 2019 IEEE 16th International Symposium on Biomedical Imaging (ISBI 2019). 2019 IEEE 16th International Symposium on Biomedical Imaging (ISBI 2019), pp. 178–181.
Wassermann D (2016) The white matter query language: a novel approach for describing human white matter anatomy, Brain Structure and Function, 221(9), pp. 4705–4721.
Zhang F (2022) ‘TractoFormer: A Novel Fiber-Level Whole Brain Tractography Analysis Framework Using Spectral Embedding and Vision Transformers’, in Wang L, Dou Q, Fletcher P T, Speidel S, and Li S (eds.) Medical Image Computing and Computer Assisted Intervention – MICCAI 2022. Cham: Springer Nature Switzerland, pp. 196–206.
Acknowledgements: ANID, Chile: Beca Doctorado Nacional 2022-21210468, FONDECYT 1221665, ANILLO ACT210053, Basal FB0008, and Basal FB210017.