Diffusion-/perfusion-weighted imaging fusion to automatically identify stroke within 4.5 h

Study population

Data from Nanjing First Hospital and the Affiliated Jiangning Hospital of Nanjing Medical University from January 2017 to December 2020 were retrospectively included. Anterior circulation acute ischemic stroke patients (AIS) were included if their symptom onset time was clear and within 24 h and if they underwent MRI scans, including both DWI and PWI sequences. Of 520 patients considered candidates for analysis, 32 patients with severely artifacts DWI or PWI images and 55 patients with lesions < 1 cc in size were excluded. Finally, a total of 433 patients were included for analysis. The flowchart shown in Fig. 1 outlines the patient selection process. According to the Trial of ORG 10172 in Acute Stroke Treatment (TOAST) classification system [19], AIS was classified into the following five subgroups: (1) large artery atherosclerosis (LAA); (2) small-artery occlusion (SAO); (3) cardioembolism (CE); (4) other determined cause (OC); and (5) undetermined cause (UND). The hospital review board of Nanjing Medical University approved the study protocol. All patients in this study provided written informed consent before the MRI examination. The patients were divided into two classes according to the onset time: positive (≤ 4.5 h) and negative (> 4.5 h).

Fig. 1figure 1

Flowchart of the included studies. DWI = diffusion-weighted imaging, DP fusion = diffusion/perfusion-weighted imaging fusion, PWI = perfusion-weighted imaging, VOI = volume of interest

MRI protocol and processing

Patients in the two centers were scanned with the same MRI scanner and parameters. MRI scans were performed on a 3.0-T MRI scanner (Ingenia, Philips Healthcare) with an 8-channel receiver array head coil. A detailed description of the MRI protocol is provided in the Supplemental Materials (online). The PWI data were analyzed using RAPID software (IschemaView 5.0.2) to obtain Tmax images. The PWI processing component consists of motion correction and adjustments for different acquisition times in multislice echo-planar imaging scans, conversion of measured MR signals to estimated changes in transverse relaxivity, automatic detection of arterial input function and venous output, correction for the nonlinear effects of gadolinium tracers in bulk blood and capillaries, deconvolution, and final computation of perfusion parameters [20]. The Tmax parameter is a bolus-shape-independent estimate of time delay for blood delivery between a main feeding artery (e.g., middle cerebral artery) and tissue at a given spatial location, and the hypoperfusion is identified based on Tmax prolongation beyond a pre-specified threshold.

Manual segmentation

All DWI and Tmax images of acute stroke patients were derived from DICOM format and converted to NII format using MRIcron software (https://www.nitrc.org/projects/mricron). The high-intensity signal infarction area on DWI (apparent diffusion coefficient < 600 × 10–6 mm2/s) and abnormal perfusion areas (Tmax > 6 s) on Tmax images were manually drawn as volumes of interest (VOIs) [20] using ITK-SNAP software (http://www.itksnap.org/pmwiki/pmwiki.php). Two certified neuroradiologists determined the VOI in consensus (Y-C.C., 8 years of experience, and M-Y.P., 15 years of experience).

Coregistration and penumbra VOI generation

The DWI matrix size was 256 × 256 × 18, and the Tmax matrix size was 128 × 128 × 25. The ischemic penumbra cannot be obtained directly by subtracting the VOI of lesions in images of different scales and layers. Therefore, an image registration method was adopted in combination with affine transformation, and the mutual information was used as the optimization criterion. Advanced normalization tools (ANTs) were used to register DWI and Tmax sequences. After registration, the DWI images were essentially consistent with the Tmax images in terms of spatial position, and the image matrix size was changed from 256 × 256 × 18 to 128 × 128 × 25. The penumbra VOI was calculated using the following formula:

$$VOI\left(penumbra\right)=VOI\left(Tmax>6s\right)-VOI\left(DWI\right),$$

Of the 433 patients, 15 patients were further excluded because of coregistration error.

DP fusion

The DP fusion image was obtained by fusing the DWI and Tmax images after registration. The image fusion method is shown in the following formula:

$$_=\frac_-_}_}+\frac_-_}_},$$

Where Idwi and Itmax are the DWI and Tmax images after registration, IFuse is the fused image, and μ and σ are the mean and variance of the image, respectively. The DP fusion images contain all imaging features of the DWI and Tmax images.

Imaging feature extraction and selection

The radiomic features of the VOIDWI, VOITmax, VOIpenumbra, and VOIfusion were computed using PyRadiomics software (version: 3.0.1, https://pyradiomics.readthedocs.io/en/latest/), which follows the image biomarker standardization initiative (IBSI). The radiomics features covered six categories: shape-based (3D) features; first-order statistical features; gray-level cooccurrence matrix (GLCM); gray-level run-length matrix (GLRLM); gray-level size-zone matrix (GLSLM); gray-level dependence matrix (GLDM). Finally, a total of 1046 features were extracted from each VOI. The mean and standard deviation of features were normalized using the Z-score method. To filter redundant features and reduce feature dimensions, the t test was first used to identify features that could significantly differentiate between the onset time groups (p < 0.05). Then, the least absolute shrinkage and selection operator (LASSO) method with tenfold cross-validation, a suitable method for high-dimensional data regression, was used to select the most useful predictive features.

Machine learning model

Two common ML algorithms were used to develop the classifier models: a support vector machine (SVM) [21] and logistic regression (LR) [22]. All training processes were performed in R software with the caret package. The models were evaluated using fivefold cross-validation. The 418 patients remaining after the previous exclusion process were divided into training and test sets at a ratio of 4:1. That is, 335 patients in each fold were included in the training set, and 83 patients were included in the test set.

Deep learning for lesion segmentation and classification

Segmentation and classification modules for identifying the onset time were designed in the same network frame, which we called X-Net, according to the shape of the model. The overall network architecture is shown in Fig. 2. The model consists of three components: a double distillation fusion encoder (differential distillation module and feature fusion module (Figure S1)), a multioutput separation decoding module (Figure S2), and a fully connected classifier. A detailed description of the model is provided in the Supplemental Materials (online). Additionally, conventional networks, including 2D Unet, 3D Unet, Vnet, and Attention Unet, were also used to segment DWI and Tmax images and compare them with X-Net. A few minutes are typically required to segment and classify a single patient with these methods. The classification and segmentation framework proposed in this study is shown in Fig. 3.

Fig. 2figure 2

Schematic of the X-Net architecture of the segmentation-classification model. The model has three components: a double distillation fusion encoder, a multioutput separation decoder, and a fully connected classifier. The encoding part includes a 3D convolution layer and a pooling layer. The encoder has two different paths for extracting features from DW images and Tmax images. The decoding part generates two outputs: Mask (DWI) and Mask (Tmax). In the multioutput separation refinement module, the output is separated and refined step-by-step to obtain the final output result. Then, the fusion abstraction feature of the last layer of the encoder is used to generate binary classification results. This component includes a flattening operation, global average pooling, fully connected layers, and a sigmoid output function

Fig. 3figure 3

Classification and segmentation framework for identifying stroke onset time using DW and PW images proposed in this study

Statistical analysis

Statistical analyses were performed using the statistical software R Studio (version 4.0.3). The Kolmogorov–Smirnov statistical test was used to test the normality of continuous variables. Continuous variables are presented as medians (interquartile ranges) and were assessed by Student’s t tests and Mann–Whitney U tests. Categorical variables are presented as percentages and were assessed by the χ2 test. Receiver operating characteristic (ROC) curve analysis, area under the curve (AUC), sensitivity, specificity, and accuracy were calculated using the pROC package to compare the efficacy of each model. Decision curve analysis (DCA) was conducted to assess the utility of each model. Calibration curves were used to evaluate whether the predicted probability of the classification model was close to the real probability. The Dice coefficient, Jaccard coefficient, average surface distance (ASD), and 95% Hausdorff (HD_95) metric were calculated to evaluate the segmentation efficacy of each model. Ablation experiments were conducted by removing one or more modules, including the differential distillation, feature fusion, and multioutput separation decoding modules, to compare the segmentation-classification efficacy of the X-Net. The test set was further divided into nine groups according to the onset time, namely, 0–1 h, 1–2 h, 3–4 h, 4–5 h, 5–6 h, 6–7 h, 7–8 h, and greater than 8 h, to assess classification efficacy for the subgroups. All statistical tests were two-sided, and p values of less than 0.05 were deemed to indicate statistical significance.

留言 (0)

沒有登入
gif