VOXEL-TO-BLADDER FULLNESS SENSATION

Abstract

Current medical diagnosis and treatment methods for neurogenic lower urinary tract dysfunction (NLUTD) disorders are constrained by our limited understanding of how a set of the complex neural circuits that regulate the LUT function. Identifying robust biomarkers for perceived bladder sensation could be key to advancing diagnostic and therapeutic modalities for NLUTD. In this work, we applied a transfer learning approach to infer bladder fullness sensation from functional magnetic resonance imaging (fMRI) data. While the proposed approach effectively represented fMRI scans in the embedding space, it did not predict bladder fullness sensation significantly better than random chance.

1. INTRODUCTION

The lower urinary tract (LUT) consists of the bladder and urethra, and its main function is storage and periodic release (voiding) of urine. Despite the seemingly simple function, the LUT is controlled by a complex set of brain regions, including the cortex, brainstem, and potentially the cerebellum.[1] The leading theory on brain-LUT control is summarized in a working model that delineates the roles of key neural circuits and the corresponding brain regions involved.[2] However, significant gaps remain in our understanding of brain control over LUT, such as the mechanisms underlying dysfunction in neurological diseases such as Multiple sclerosis and Parkinson’s disease, as well as the neural dynamics of bladder control in healthy subjects.[3,4] The goal of this work was to train a deep learning model on functional magnetic resonance imaging (fMRI) scans obtained from healthy subjects, to shed light on the inner workings of brain control of the LUT.

Previously, deep learning models have demonstrated better performance than random chance for speech[5] and vision reconstruction tasks, [6,7] where the objective was to infer the imagined or perceived speech/vision from the activations captured in simultaneously recorded fMRI scans. A comparable task for LUT control would be to infer the bladder fullness level from fMRI data. The most recent related work on LUT control applies machine learning approaches to predetermined features of fMRI data, such as functional connectivity in set regions of interest (ROIs) during resting-state scans.[8] Nevertheless, neurons crucial for the LUT control task are not necessarily confined within predefined spherical ROIs but rather in a topology dictated by the brain anatomy of the respective LUT control regions, such as the periaqueductal gray (PAG). Since the deep learning models, such as those utilized with the aforementioned speech and vision reconstruction tasks, train on raw fMRI data rather than a predetermined set of ROIs, they are more promising in theory. To our knowledge, no similar deep-learning approach has been attempted for the LUT control task.

Here, we employ a deep learning model for resolving bladder fullness sensation from acquired fMRI blood oxygen level-dependent (BOLD) activations. Specifically, a representation of BOLD signals in the entire cortical and subcortical areas was sought as it pertains to bladder fullness sensation. If successful, this representation and all dynamics of the respective region recruitment can be attributed to overactive or underactive bladder patient scans to study the variations in activation patterns between pathologies. In addition, such low dimensional representations might explain the coordination of related brain regions, which are not explained by the leading theory on supraspinal LUT control.[1,2]

2. MATERIALS AND METHODS

2.1. Objective

The data modality that is currently primarily used for brain control over LUT research is fMRI[2], which essentially is 4D scans of the whole brain over time, X_t, usually sampled about 1–3 second rate. The objective of this project is to infer the subject’s bladder fill level sensation y_t ∈[0,1], where 0 is the lowest perceivable sensation, and 1 is the highest or “full” sensation with a great urge to void, given the subject’s fMRI snapshot at that point in time $X_t\in {ℝ}^{d_x\times d_y\times d_z}$, which has three spatial dimensions.

2.2. Data

Firstly, fMRI scans, X_t, lie in a high dimensional space where activations of various brain regions are associated with specific tasks.

For this analysis, a task-specific dataset was obtained from a previous study completed by our group (registered at ClinicalTrials.gov NCT04846387 [9]), herein referred to as the “primary dataset”.[9] In summary, it contains the fMRI scans of 19 healthy subjects (1 out of 20 is missing timing labels) whose bladders are naturally filling while they are in the scanner. The protocol is designed so that participants start with an empty bladder state, y₀ = 0. Once the participant indicates a “full bladder” sensation y_T = 1, additional holding and voiding (or attempted voiding) tasks are performed, but those sessions are omitted in this work.

The primary dataset is tailored to the task. However, its main limitation is that there are 19 subjects, each of which has about 600 samples on average. This dataset has the desired labels y_t; however, the sample size is too small to train an interesting neural network. For this reason, a publicly available dataset is used, referred to herein as the “secondary dataset”. The secondary dataset is a bigger fMRI dataset, though not specific to the LUT control task. It is used to learn meaningful low-dimensional representations of the fMRI scans, which are transferred to a shallow neural network that trains on the primary dataset. The criteria used to pick the secondary dataset included (1) containing high-definition 7-Tesla fMRI scans and (2) having large amounts of subject recordings, though not necessarily LUT control task-related. After some literature research, the Natural Scenes Dataset has been chosen as the secondary dataset, which coincidentally is currently the largest high-definition task-based fMRI dataset of any tasks publicly available, containing more than 300 hours of recordings.[10]

For simplicity, the objective stated in section 2.1 is broken down into two steps. “Step 1” is learning meaningful dimensional representations of snapshots the fMRI scans X_t, and the secondary dataset is reserved for this step. “Step2” is training a model that maps the learned representation in step 1 to the desired bladder fullness sensation variable y_t. For step 1, an autoencoder is trained, and for step 2, the trained encoder is used as a feature extractor, after which a shallow neural network is trained for task-specific inference (Fig. 2).

FIGURE 2:

(a) The autoencoder is trained on the secondary dataset. (b) The encoder from the trained autoencoder is transferred, and the shallow neural network is trained on the primary dataset.

2.3. Preprocessing

The conventional preprocessing pipeline for fMRI data modality includes steps such as slice timing correction, co-registration with anatomical scans, motion correction, skull stripping, and normalization to MNI coordinates.[11,12] These steps are crucial for reducing noise and artifacts due to factors not related to brain region BOLD activations. These standard fMRI preprocessing steps for both datasets were employed. The secondary dataset included precomputed derivatives of these preprocessing steps[10] via the “fmriprep” tool[12]. The derivatives do not include spatial filtering due to variability in the application, and no temporal filtering was done for either step.

2.4. Data Wrangling

Even after the preprocessing is done on the primary and secondary fMRI datasets, some adjustments must be made due to the relatively large size of imaging sessions and due to the choice of neural network architecture. The skull extraction step yields compressed fMRI scans that are of various shapes. So first, the input to the model is determined based on the maximum size of cropped scans and down-sampled in shape to fit the stacks of convolutional layers for convenience. Secondly, the compressed “NIFTI” files corresponding to the scans are processed, divided into batches, down-sampled in numerical precision to conserve computational resources, serialized, and saved into the memory via a format that can be read faster during training. The choice of batch size turned out to be a crucial parameter affecting training speed and model performance.

2.5. Model Architectures

For step 1, the goal was to leverage the local interaction of the brain regions to represent the whole scan. For this purpose, the encoder architecture consists of three 3D convolutional and max pooling layers (Fig. 3). The input lies in the 3D voxel space, and the encoder will give lower-dimensional embeddings of each scan snapshot. The decoder architecture is obtained by reversing the order of these layers and replacing convolutional layers with transposed convolutional layers of the same size. The respective autoencoder is trained by minimizing the mean squared reconstruction error on the secondary dataset. Slightly different neural network model architectures of encoders/decoders were tested with different filter sizes and a dense layer following the last max pooling layer. The architecture with the best reconstruction error was chosen which is given in Figure 3.

FIGURE 3:

The Encoder architecture that is trained in step 1 and used in step 2.

For step 2, the shallow network is chosen as 3 dense layers with 64, 16, and 16 neurons. Mean squared logarithmic error was used as the loss. Success for the overall objective is judged by loss significantly lower than random chance inference, which was the standard for the aforementioned image reconstruction and speech perception tasks.[5,6]

3. RESULTS AND DISCUSSION

3.1. Step 1 Results

For the proposed architecture and batch size of 30 scans, the original and reconstructed scans for a slice along the z-axis are given in Figure 4. The reconstructed scans qualitatively resemble the original ones.

FIGURE 4:

A slice of the same fMRI scan with the original scan on the left and reconstructed scan by the autoencoder on the right.

3.2. Step 2 Results

The prediction step on the validation set where red points are individual scans, and for every batch, their target value are close to each other since they are taken from a consecutive run (Fig. 5).

FIGURE 5:

Prediction on validation set. The red points are unseen ground truth bladder fullness sensations, and the blue line represents the predicted value. The model has fitted to the mean of the sample distribution.

3.3. Discussion

As more public fMRI datasets are becoming available, we are seeing more and more comprehensive brain-computer interface models that aim to decode the brain. As mentioned, for vision[6,7] and speech[5] tasks, previous work has achieved better than random chance accuracy for reconstructing perceived image/audio. However, no such attempts have been found in the field of urology, where a full working theory is still being developed.[1–3] Consequently, this was the first-ever known attempt at reconstructing bladder fullness sensation from the fMRI modality.

Regarding the results achieved, after extensive training, step 1 was solved with reasonable performance, as indicated in Figure 4. However, even after testing with different batch sizes, architecture, and other learning hyperparameters, step 2 is not solved with a better than random choice performance, as indicated by Figure 5. The model does not do better than fitting to the mean of the training data distribution.

Several possibilities explain the result obtained in step 2. The biggest identified challenge toward this objective has been the confounding. For the image and speech reconstruction problems, the related regions in the cortex primarily process speech or visual stimuli. However, in LUT control, many of the regions are responsible for non-LUT control-related functions, such as the PAG, which regulates breathing and cardiovascular homeostasis. Consequently, the respective BOLD activations are not a clear indicator of bladder fullness sensation. The expectation here was that given the meaningful representation of the high dimensional fMRI data, the model would be able to find patterns that correlate directly with the sensation that avoids all confounders. This was not the case. An improvement for this could be a modification in the design so that representations learned in step 1 are particular to LUT control. However, such a solution must incorporate labeled and unlabeled samples into the optimization problem in training. Another remedy for the confounding problem is to make use of prior knowledge of the LUT control. Even though the working model is not absolute, it is the result of years of experiments that align.[2] Making use of this prior knowledge can limit the noise due to confounding variables.

Another merit of the fMRI data modality that is omitted in this work, which might have significant implications in step 2 performance, is the temporal dynamics of BOLD signals. The idea is that the region activations are not at a given point in time, but the sequence of activations carries the necessary information about the target variable. It is possible to add attention mechanisms to the design, which should capture the context due to previous activations.

4. CONCLUSION

In this work, a deep learning model was proposed that attempts to decode brain region BOLD activations with respect to bladder fullness sensation in the LUT. The proposed problem is novel and potentially has direct implications in clinical urology as well as basic science. The problem was broken down into two steps, namely, low dimensional embeddings of fMRI data and mapping to the target variable. Step 1 was successfully solved, but step 2 was not by the proposed approach. Confounding and stationarity assumptions were identified as the most critical limitations of this approach.

FIGURE 1:

The objective of this work is to predict “bladder fullness sensation” from observed fMRI BOLD activations.