DeLTA-BIT: an open-source probabilistic tractography-based deep learning framework for thalamic targeting in functional neurological disorders

TL;DR

This work addresses the need for fast, patient-specific thalamic targeting in functional neurological disorder treatments by proposing DeLTA-BIT, a open-source CNN-based framework that predicts VIM location from $T1$ MRI while leveraging probabilistic tractography-derived targets. Trained on the HCP dataset, the model is validated internally against tractography-derived ROIs and externally against clinical data, achieving a mean DSC of $0.62 \pm 0.15$ and sDSC of $0.76 \pm 0.17$ while predicting in fractions of a second per subject. External validation shows comparable performance to atlas-based methods with substantially faster inference, though GT variability and dataset differences introduce challenges. Overall, DeLTA-BIT offers real-time VIM localization potential for MR-guided procedures and serves as an open platform for further enhancements with clinical data.

Abstract

In the last years in-vivo tractography has assumed an important role in neurosciences, for both research and clinical applications such as non-invasive investigation of brain connectivity and presurgical planning in neurosurgery. In more recent years there has been a growing interest in the applications of diffusion tractography for target identification in functional neurological disorders for an increasingly tailored approach. The growing diffusion of well-established neurosurgical procedures, such as deep brain stimulation or trans-cranial Magnetic Resonance-guided Focused Ultrasound, favored this trend. Tractography can indeed provide more accurate, patient-specific, information about the targeted region if compared to stereotactic atlases. On the other hand, this tractography-based approach is not very physician-friendly, and its heavily time consuming since it needs several hours for Magnetic Resonance Imaging data processing. In this study we propose a novel open-source deep learning framework called DeLTA-BIT (acronym of Deep-learning Local TrActography for BraIn Targeting) for fast target predictions, based on probabilistic tractography. The proposed framework exploits a convolutional neural network (CNN) to predict the location of the Ventral Intermediate Nucleus of the thalamus (VIM). The CNN was trained on the Human Connectome Project (HCP) dataset. The model capability in predicting the VIM location was tested both on the HCP (internal validation) and clinical data (external validation). Results from the internal validation have shown good capability in predicting the VIM region (mean DSC = 0.62+- 0.15, mean sDSC=0.76+- 0.17) by using just T1 images as input, in a time scale of fraction of second per subject. As for the clinical data, results have been compared with an atlas-based method demonstrating similar performance, but within a significantly shorter timeframe.

Figures (11)

• Figure 1: Schematic representation of the probabilistic tractography pipeline.
• Figure 2: Probabilistic tractography streamlines for the DRTC tract. In blue the left-M1-cortex and the right dentate nucleus ROIs. In white the left thalamus. In red-to-yellow scale the DRTC tractogram obtained by merging the dentate-thalamus-M1 tractography in both directions. Panel a, from left to right: sagital, coronal and axial projections. Panel b: rendering of the 3D image.
• Figure 3: M1 connectivity map obtained setting the left thalamus as seed and the M1 as target region. As is well known, the VIM is located within this area.
• Figure 4: Identification of the VIM region. The VIM ROI was obtained by applying a mean filter (kernel size $= 5\times5\times5$) to the raw distribution obtained by intersection. The contour of the obtained binary mask is highlighted in blue.
• Figure 5: Schematic representation of the U-Net neural network. This network exhibits a symmetric shape, working as an encoder-decoder: the left side corresponds to the encoder, while the right side represents the decoder. The network is organized into levels, with each level containing two convolutional blocks. Each convolutional block consists of a 3D convolutional layer (kernel size 3), followed by a LeakyReLU activation and Batch Normalization layers. The numerical values within the blocks (in parentheses) denote the block outputs shape. Specifically, the first three numbers indicate spatial dimensions, while the last number represents the number of feature maps. Additionally, the numbers at the bottom of the blocks correspond to the convolutional stride. A stride of 1 maintains the same spatial size for the layer output, while a stride of 2 spatially reduces the output by half. At each level, the initial convolutional block employs a stride of 2 (except for the first level), effectively halving the input size in all directions. This reduction in spatial dimensions is compensated by doubling the number of feature maps. The output layer consists of a convolutional block with a sigmoid activation function, yielding output values within the range 0-1, representing the probability for a voxel to belong to the VIM region.