Automatic assessment of disproportionately enlarged subarachnoid-space hydrocephalus from 3D MRI using two deep learning models

Abstract

Background: Disproportionately enlarged subarachnoid-space hydrocephalus (DESH) is a key feature for Hakim disease (idiopathic normal pressure hydrocephalus: iNPH), but subjectively evaluated. To develop automatic quantitative assessment of DESH with automatic segmentation using combined deep learning models.

Methods: This study included 180 participants (42 Hakim patients, 138 healthy volunteers; 78 males, 102 females). Overall, 159 three-dimensional (3D) T1-weighted and 180 T2-weighted MRIs were included. As a semantic segmentation, 3D MRIs were automatically segmented in the total ventricles, total subarachnoid space (SAS), high-convexity SAS, and Sylvian fissure and basal cistern on the 3D U-Net model. As an image classification, DESH, ventricular dilatation (VD), tightened sulci in the high convexities (THC), and Sylvian fissure dilatation (SFD) were automatically assessed on the multimodal convolutional neural network (CNN) model. For both deep learning models, 110 T1- and 130 T2-weighted MRIs were used for training, 30 T1- and 30 T2-weighted MRIs for internal validation, and the remaining 19 T1- and 20 T2-weighted MRIs for external validation. Dice score was calculated as (overlapping area) × 2/total area.

Results: Automatic region extraction from 3D T1- and T2-weighted MRI was accurate for the total ventricles (mean Dice scores: 0.85 and 0.83), Sylvian fissure and basal cistern (0.70 and 0.69), and high-convexity SAS (0.68 and 0.60), respectively. Automatic determination of DESH, VD, THC, and SFD from the segmented regions on the multimodal CNN model was sufficiently reliable; all of the mean softmax probability scores were exceeded by 0.95. All of the areas under the receiver-operating characteristic curves of the DESH, Venthi, and Sylhi indexes calculated by the segmented regions for detecting DESH were exceeded by 0.97.

Conclusion: Using 3D U-Net and a multimodal CNN, DESH was automatically detected with automatically segmented regions from 3D MRIs. Our developed diagnostic support tool can improve the precision of Hakim disease (iNPH) diagnosis.

Keywords: MRI; artificial intelligence; chronic hydrocephalus in adults; deep learning; disproportionately enlarged subarachnoid-space hydrocephalus; hakim disease; idiopathic normal pressure hydrocephalus; tightened sulci in high convexities.

Figure 1

Segmentation from three-dimensional T1- and T2-weighted MRI. The upper axial (A), sagittal (B), and coronal (C) images on 3D T1-weighted MRI show the results for fully automatically segmented regions, including total ventricles (green) and total subarachnoid spaces (marine blue) of a representative patient with Hakim disease and DESH, using the Brain Subregion Analysis application on the 3D volume analyzer SYNAPSE 3D workstation (FUJIFILM Corporation). The lower three-dimensional (D), sagittal (E), and coronal (F) images on 3D T2-weighted MRI show the results of manually segmented total ventricles (light blue in D) and total subarachnoid spaces (light green in E,F) of a representative healthy elderly volunteer.

Figure 2

Input image masks as the ground truth labels transferred to the cloud-based AI development service. The upper axial (A), sagittal (B), and coronal (C) images on 3D T1-weighted MRI in the same Hakim patient as the upper panel of Figure 1 show the input image masks including total ventricles (green), Sylvian fissure and basal cistern (purple), high-convexity part of the subarachnoid space (yellow), and the other subarachnoid spaces (marine blue). The lower axial (D), sagittal (E), and coronal (F) images on 3D T2-weighted MRI in the same healthy volunteer as the lower panel of Figure 1 show the input image masks for deep learning including total ventricles (light green), Sylvian fissure and basal cistern (pink), high-convexity part of the subarachnoid space (yellow), and the other subarachnoid spaces (light blue).

Figure 3

Two combined deep learning models; and multimodal convolutional neural network for image classification. (A) 3D U-Net model with four layers for volumetric semantic segmentation task. Each blue box corresponds to a multi-channel feature map. The number of channels is denoted on front of the box. White boxes indicate copied feature maps. The color arrows indicate each process: light blue arrows indicate convolution (Conv) with kernel size (3, 3, 3) in addition to batch normalization (BN) and rectified linear unit (ReLU) activation layer; red arrows indicate max-pooling with kernel size (2, 2, 2); green arrows indicate up-convolution (Up-Conv) with kernel size (3, 3, 3) and dilation rate (2, 2, 2) in addition to BN and ReLU; and gray arrows indicate direct concatenation from each encoding layer of feature map extracted by downsampling to the corresponding decoding layer of feature map by upsampling. Signal values were normalized by percentile (minimum 0.05, maximum 0.95) as a preprocessing step. (B) Multimodal convolutional neural network for image classification task. Each blue box corresponds to a multi-channel feature map with the number of channels denoted on the front of the box. The color arrows indicate each process: purple arrows indicate convolution (Conv) with kernel size (3, 3, 3) in addition to batch normalization (BN), rectified linear unit (ReLU) activation, self-attention, and pooling layer; turquoise blue arrows indicate global average pooling (GAP) or fully connection (FC) layer. In the embedding layer, all input variables were transformed into the feature maps. At the end of the last convolutional layer, the final feature maps were fed to a softmax activation function to generate a probability score for each class. The image intensities of input images were normalized to [0, 1] by their maximum and minimize values.

Figure 4

Results of training for deep learning of the semantic segmentation. The Dice scores (emerald green line), loss (lime green line), precision (blue), and recall (purple) for the automatically segmented volumes of the total ventricles (A,B), total subarachnoid spaces (C,D), high-convexity part of the subarachnoid space (E,F), and Sylvian fissure and basal cistern (G,H) on 3D T1-weighted (A,C,E,G) and T2-weighted MRIs (B,D,F,H).

Figure 5

Inference results for internal validation of the semantic segmentation and image classification. The Dice scores (emerald green line), loss (lime green line), precision (blue), and recall (purple) for the automatically segmented volumes of the total ventricles (A,B), total subarachnoid spaces (C,D), high-convexity part of the subarachnoid space (E,F), and Sylvian fissure and basal cistern (G,H) on 3D T1-weighted (A,C,E,G) and T2-weighted MRIs (B,D,F,H).

Figure 6

Results of training for deep learning of the image classification. The accuracy (blue line) and loss (green line) for the detection of disproportionately enlarged subarachnoid space hydrocephalus (DESH: A,B), ventricular dilatation (VD: C,D), tightened sulci in the high convexities (THC: E,F), and Sylvian fissure dilation (SFD: G,H) on 3D T1-weighted (A,C,E,G) and T2-weighted MRIs (B,D,F,H).

Figure 7

Inference results for internal validation of the image classification. The loss (green line) for the detection of disproportionately enlarged subarachnoid space hydrocephalus (DESH: A,B), ventricular dilatation (VD: C,D), tightened sulci in the high convexities (THC: E,F), and Sylvian fissure dilation (SFD: G,H) on 3D T1-weighted (A,C,E,G) and T2-weighted MRIs (B,D,F,H).

Figure 8

Comparison between manually and automatically segmented regions from 3D T1-weighted images. 3D T1-weighted images in a representative healthy volunteer (A–F) and a representative patient with Hakim disease and DESH (G–L): manually segmented (A–C,G–I) and automatically segmented (D–F,K–L) volumes of the total ventricles (green); Sylvian fissure and basal cistern (purple); high-convexity part of the subarachnoid space (yellow); other subarachnoid spaces (marine blue) from 3D T1-weighted images.

Figure 9

Comparison between manually and automatically segmented regions from 3D T2-weighted images. 3D T2-weighted images in a representative healthy volunteer (A–F) and a representative patient with Hakim disease and DESH (G–L): manually segmented (A–C,G–I) and automatically segmented (D–F,K–L) volumes of the total ventricles (green); Sylvian fissure and basal cistern (purple); high-convexity part of the subarachnoid space (yellow); other subarachnoid spaces (marine blue) from 3D T2-weighted images.

Figure 10

A case of discrepancy in DESH determination between AI and expert. MRI of an 84-year-old male volunteer, who claimed no specific history of head trauma showed a signal deficit (white arrow) in the left frontal region due to a metal artifact. The AI automatically judged the presence of DESH (softmax probability score: 1.0), VD (1.0), SFD (0.75), and the absence of THC (0.84), while the expert judged the presence of VD but not DESH, THC, or SFD. This case should have been excluded from the study.