Identifying Patients with CSF-Venous Fistula Using Brain MRI: A Deep Learning Approach

Patient Cohort

A retrospective analysis was carried out, focusing on patients suspected of having SIH who underwent lateral decubitus DSM between December 30, 2021, and November 30, 2022. The exclusion criteria included the absence of pre-DSM brain MR imaging or substandard image quality. In addition, to focus on the SIH subpopulation associated with CSF-venous fistula, patients were excluded if a spinal longitudinal extradural fluid collection (SLEC) was noted on the pre-DSM spinal MR imaging. For any patients who had undergone a previous DSM (ie, before the study inclusion date), either 1) the initial DSM identified a CSF leak or 2) the first DSM performed at our institution (in patients for whom a CSF leak was not identified) and the most recent brain MR imaging preceding the DSM were used for analysis. The number of days between the pre-DSM brain MR imaging and analyzed DSM was recorded.

Digital Subtraction Myelography and CT Myelography Technique

Our previously described DSM technique⁸ spanned 2 days, with patients being positioned first in the right lateral decubitus position and subsequently in the left. DSMs were performed using biplane fluoroscopy machines, specifically, either the Artis Icono system (Siemens Healthineers) or the Allura Xper system (Philips Healthcare), with most cases using only a single anteroposterior plane. Patients were placed on a wedge-shaped cushion on a tiltable table such that their hips were higher than their shoulders. A 20- or 22-gauge spinal needle was inserted into the thecal sac, usually at the L2–3 or L3–4 level. Each side underwent 2 distinct DSM acquisitions, with one focusing on the upper spine and the other on the lower. A total of 11 mL of intrathecal Omnipaque 300 (GE Healthcare) was used. Post-DSM, the patients were transferred to a CT scanner for a full-spine same-side lateral decubitus CTM, which was conducted on a dual source Somatom Flash scanner (Siemens Healthineers). Both the DSM and the CT myelogram were analyzed as parts of a unified diagnostic assessment.

MR Brain Imaging Protocol

Brain MRIs were performed by using either 1.5T or 3T scanners. Most image analyses relied on fat-saturated postcontrast 3D T1 fast spin echo sequences (TR = 600 ms, TE = 7.2 ms, flip angle = 120°, section thickness = 1 mm, FOV = 250 × 250 mm²).

Image Analysis

Four neuroradiologists (I.T. M., A.A.M., J.C.B., and J.T.V.) with 2–8 years of posttraining subspecialty expertise in interpreting neuroradiologic examinations reviewed the DSM and pre-DSM MR brain images, with the entire cohort being split evenly among the reviewers. The reviewers were blinded to clinical information but not to the official reports or annotations in the picture archiving and communication system. Each reviewer independently evaluated all DSMs for the presence or absence of an identified CSF leak (specified as positive, negative, or indeterminate). MR imaging brain scans were assessed for multiple potential stigmata of SIH that must be measured to determine a Bern score, including the engorgement of venous sinuses, pachymeningeal enhancement, suprasellar cistern height, subdural fluid collection, prepontine cistern width, and mamillopontine distance. The Bern score was determined for each patient, and its discrimination performance was evaluated by calculating the area under the receiver operating characteristic curve (AUROC).

Data Set Splitting

Participants whose DSM studies demonstrated an absence or definitive presence of CSF-venous fistula were included in the study. The data were split into 5 folds at the patient level by using the GroupKfold module from the scikit-learn package, version 1.2.0. Cross-validation of the final model was performed on all 5 groups to determine the robustness of the results.⁹

Data Preprocessing and Model Development

Initially, for CSF-venous fistula detection, the process involved converting 2D sagittal contrast-enhanced T1-weighted (CET1) images from the DICOM format to the 3D Neuroimaging Informatics Technology Initiative format. Subsequently, the MR images were resampled to possess voxel dimensions of 1 × 1 × 1 mm through trilinear interpolation. To satisfy image size uniformity, the images were zero-padded to dimensions of 240, 260, and 260 voxels, corresponding to the largest dimensions observed across the data set.¹⁰ To prevent overfitting, a range of data augmentation techniques were applied.¹¹ These encompassed random intensity shifts, random histogram adjustments, random bias field variations, flipping, random affine transformations, and random Gaussian noise. For further descriptions of these techniques, refer to the Online Supplemental Data.

We used a 3D-DenseNet-121 classifier sourced from the MONAI package (version 1.1.0), which showed promising results in classification tasks dealing with 3D medical imaging volumes.^10,12,13 DenseNet, a convolutional neural network, uses convolutions to extract meaningful information while connecting each layer to every other layer in a feed-forward manner. Within the architecture of DenseNet, a 1 × 1 convolution was introduced before each convolution layer to serve as a bottleneck layer to trim the count of feature mappings.¹⁴ This approach to dimension reduction, implemented through both bottleneck layers and transition layers within each attenuated block, contributes to enhanced parameter efficiency and a reduction in model complexity. This design choice aids in mitigating the risk of overfitting. To enhance the model’s generalizability and mitigate overfitting, we used the AdamW optimizer with a batch size of 4 and a weight decay of 0.1.¹⁵ This choice of optimizer integrates weight decay as a regularization technique. The scheduler started with an initial learning rate of 0.001 and ran for 450 epochs, and this was followed by a restart interval of 200 epochs. This approach was adopted to circumvent potential issues with converging to local minima.¹⁶ Figure 1 illustrates the schematic pipeline for preprocessing and training.

• Download figure
• Open in new tab
• Download powerpoint

FIG 1.

Schematic illustration of the preprocessing, data augmentation, and model training pipeline. A, 3D contrast-enhanced T1 input volume. B, Preprocessed and augmented CET1 to mitigate overfitting. C, 3D-DenseNet model. D, Prediction of absence or presence of CSF leak.

To handle the slight class imbalance in the data set, we utilized a weighted cross-entropy loss function and assigned reverse class ratios as weights to the positive and negative classes to more effectively address the imbalance.¹⁷ The model weights from the epoch with the greatest AUROC, indicating the highest discriminative value, were saved for every fold.¹⁸ In addition to each fold, the mean and standard deviation of the AUROC across the folds were reported.

Occlusion Interpretation Maps

In our study, we used an occlusion interpretation technique to investigate the significance of different regions within 3D CET1 volumes for model predictions. This approach systematically replaces portions of the input volume with zeros and measures the impact on the model’s output logits to identify the most influential regions in the CET1 volume for the model’s decision-making process. A 3D patch with a shape of 12 × 13 × 13 was slid across the input CET1 volume, and the change in the model’s prediction score for the target label was recorded at each step. These changes were accumulated to form an occlusion map, which served as a visual interpretation tool. The stride for moving the patch through the input CET1 volume was set to 12 × 13 × 13, ensuring a comprehensive yet computationally efficient exploration of the 3D volume space.

Our model training was conducted on a cluster comprising 4 NVIDIA A100 GPUs. All image processing and model development were performed using PyTorch 1.12.0 and MONAI 1.1.0 with Python 3.10.4 on a GPU cluster of 4 GPUs (NVIDIA A100). All statistical analyses were performed using scikit-learn 1.2.0.