Scientific Papers

High-resolution magnetic resonance imaging-based radiomic features aid in selecting endovascular candidates among patients with cerebral venous sinus thrombosis | Thrombosis Journal

Description of Image

Patient enrollment

The ethical committee of Beijing Tiantan Hospital (KY2016-039-02) approved this study, and all patients provided written informed consent.

For our prospective CVST cohort, we retrospectively included 53 of 72 patients diagnosed with CVST who underwent HRMRI before standard drug treatment at our institution between September 2020 and September 2022 (Fig. 1). The included patients met the following criteria: (1) availability of high-quality HRMRI data before initiating medical treatment and (2) receipt of standard treatment by the Chinese Stroke Association guidelines [1]. The exclusion criteria were as follows: (1) absolute contraindication to EVT, (2) failure to achieve recanalization after endovascular mechanical thrombectomy, and (3) loss to follow-up or recurrence of thrombosis during the study period. The outcomes of patients receiving the drug and combined treatment were assessed based on main symptoms and the modified Rankin Scale (mRS) [9]. Based on the aims of this study, a good outcome was defined as achieving an mRS score ≤ 2 and experiencing [10] relief from the main symptoms. All patients were followed up for 1 month after discharge.

Fig. 1
figure 1

Flowchart with the inclusion and exclusion criteria: 72 patients diagnosed as CVST underwent HRMRI from September 2020 to September 2022. 19 patients were excluded following exclusion criteria. Two experienced neuroradiologists independently draw the RIO from the HRMRI sequences. All 106 segmentations were randomly separated into the training group and the validation group for further analysis

CVST treatment

Patients diagnosed as having CVST were admitted, and all of them received drug treatment. Low-molecular-weight heparin (LMWH) or nadroparin (0.6 ml) was subcutaneously injected twice a day at treatment initiation, with treatment lasting for 1–4 weeks. Oral administration of warfarin was used to maintain an international normalized ratio of 2–3 that lasted for 3–6 months. Patients with concurrent intracranial hemorrhage were treated with rivaroxaban (15–20 mg daily) instead of warfarin. Failure of conservative treatment was defined as the absence of relief or deterioration of symptoms after 3–5 days of drug treatment. These patients underwent EVT for further treatment.

For EVT, we selected the aspiration-first strategy for all patients. An 8-F guiding catheter was placed at the jugular bulb level, and a 6-F intermediate catheter (132 cm Catalyst, Stryker, Fremont, CA) was navigated to the superior sagittal sinus (SSS) over a 260-cm glidewire, which macerated the thrombosis. Next, we switched to a 300-cm 0.014-inch Command microwire (Abbott, Chicago, IL, USA) and introduced a 4-mm × 30-mm-sized balloon into the SSS. Using a combination of balloon dilation and aspiration via the Catalyst, we could remove the thrombosis in a distal-to-proximal direction. After thrombectomy, a 0.027-inch microcatheter was left in the SSS to facilitate continuous delivery of urokinase for 3–5 days at a dose of 25,000–50,000 IU/h.

Image acquisition

All images were acquired using a 3.0-Tesla system (Ingenia CX, Philips Healthcare, Best, The Netherlands) equipped with a 32-channel head coil. T1-weighted-MSDE, T2-weighted-MSDE, T1-weighted-contrast-MSDE, and T1-weighted-contrast images were used to delineate the lesion and extract the location and RFs. T1-weighted-MSDE image parameters were as follows: repetition time, 800 ms; echo time, 21.081 ms; flip angle, 90°; 228 slices; voxel size, 0.681 × 0.681 × 0.7 mm3; acquisition matrix, 344 × 343; scanning technique, TSE; and pixel bandwidth, 349 Hz/pixel. T2-weighted-MSDE parameters were as follows: repetition time 2,500 ms; echo time, 190.358 ms; flip angle, 90°; 229 slices; voxel size, 0.681 × 0.681 × 0.7 mm3; acquisition matrix, 344 × 342; scanning technique, TSE; and pixel bandwidth, 349 Hz/pixel. T1-weighted-contrast-MSDE image parameters were as follows: repetition time, 800 ms; echo time, 21.453 ms; flip angle, 90°; 228 slices; voxel size 0.681 × 0.681 × 0.7 mm3; acquisition matrix, 344 × 343; scanning technique, TSE; and pixel bandwidth, 349 Hz/pixel. T1-weighted-contrast image parameters were as follows: repetition time, 6.526 ms; echo time, 2.998 ms; flip angle, 8°; 196 slices; voxel size, 1.0 × 1.0 × 1.0 mm3; acquisition matrix, 240 × 240; scanning technique, TSE; and pixel bandwidth, 241 Hz/pixel.

RF extraction

Segmentation of the region of interest (ROI) started from the anterior third of the SSS, extending along the transverse sinus and sigmoid sinus with thrombosis, to the end of the internal jugular vein. Two experienced neuroradiologists who were blinded to each other used the open-source software 3D Slicer ( to delineate ROIs on the 3D T1-contrast image. To include features in other sequences, the T1-weighted-MSDE, T2-weighted-MSDE, and T1-contrast-MSDE images were co-registered to the 3D T1-contrast image with SPM12, and all voxels of the images were resampled to 1.0 × 1.0 × 1.0 mm3. A total of 1,274 RFs were extracted from each sequence using the “PyRadiomics” package in Python (3.6.4) [11]. Among the RFs, there were 18 first-order statistic features, 22 Gy-level co-occurrence matrix texture features, 16 Gy-level run-length matrix texture features, 16 Gy-level size zone matrix (GLSZM) texture features, 14 Gy-level dependence matrix texture features, and 5 neighboring gray-tone difference matrix features. In total, there were 91 original filter features, 455 Laplacian of Gaussian filter image features, and 728 wavelet filter features.

Construction of the RF signature model

RFs were used to construct a signature model (RF signature model). To increase the sample size, we combined all segmentations drawn by the neuroradiologists. All segmentations were randomly separated into training and validation groups at a ratio of 6:4. In the training group, the least absolute shrinkage and selection operator (LASSO) with three-fold cross-validation was used to select informative features [12]. Then, we calculated the RF signature as follows:

RF signature = ∑feature values × Cox efficient of feature.

The performance of the RF signature in discriminating EVT was tested in the training and validation datasets using sensitivity, specificity, accuracy, and F1-score. The optimal cutoff point was determined using the Youden index. Finally, the relationship between the RF signature model and clinical characteristics was also examined.

Construction of the support vector machine (SVM) model

The features extracted from the LASSO regression were further used to build an SVM model (CVST-SVM). A grid search method was performed to tune the hyperparameters of the SVM. The gamma value and cost parameter C were tested at 0.0001, 0.001, 0.01, 0.1, 1, 10, and 100 [13]. The prediction accuracy was maximized using the best combination of gamma and parameter C. To further evaluate the CVST-SVM model, the mean variable importance of over 1,000 permutations was used to rank the explanatory importance of the included RFs.

Statistical analyses

Statistical analyses were performed using R v3.6.1 (R Foundation for Statistical Computing, Vienna, Austria), SPSS 26.0 (SPSS, Inc., Chicago, IL, USA), and Prism 8 (GraphPad Software, Inc., La Jolla, CA, USA). The chi-square test was performed to assess the distribution of clinical characteristics between the two treatment groups. Two-tailed Student’s t-test was performed to compare risk scores in patients grouped on the basis of other clinical characteristics or stratified by risk score. Moreover, two-tailed Student’s t-test was performed to evaluate the top three features in the CVST-SVM model in the two treatment groups. P-values < 0.05 were considered statistically significant in all analyses.

Description of Image

Source link