Welcome to the WSC 2021 Interactive Program

Background and Aims

Intracranial vessel perforation is a procedural complication during endovascular thrombectomy (EVT). Its occurrence is strongly associated with unfavorable treatment outcomes. Early identification of perforation would allow therapeutic actions to prevent situation deterioration. Due to its low occurrence and the large heterogeneity in image appearance, perforations may initially be missed by the interventionalists. In this work, we study the feasibility of automated intracranial vessel perforation detection during EVT using deep learning techniques.

Methods

From the MR CLEAN registry, fifty-three patients (149 acquisitions) with vessel perforations were identified and annotated by an experienced neuroradiologist. Another 150 acquisitions (from 150 patients) without perforations were randomly selected as negative samples. The proposed solution builds on top of state-of-the-art object detection algorithms. It incorporates temporal information of DSA with bidirectional convolutional gated recurrent units (Bi-ConvGRU), further followed by a problem-tailored acquisition level optimization to reduce false positives based on temporal consistency.

Results

In ten-fold cross-validation on 203 patients (299 acquisitions, 3607 images), the proposed method achieves an area under the receiver-operating characteristic curve of 90% for acquisition-level classification of acquisitions with a perforation. The sensitivity and specificity were 81% and 86%, respectively.

Conclusions

The proposed deep learning-based algorithm achieves promising performance in vessel perforation detection in DSA for stroke patients, and can potentially be deployed in clinical practice to detect perforation early and allow direct clinical decision making.

Background and Aims

Fast identification of large vessel occlusion (LVO) at primary stroke centres (PSC) and timely referral to a comprehensive centre (CSC) are critical steps to improve outcomes from mechanical thrombectomy (MT). Increasingly artificial intelligence (AI) decision aid tools are deployed to facilitate rapid identification of LVO. In our PSC we incorporated e-Stroke software (Brainomix, Oxford, UK) into the hyperacute stroke pathway. We evaluated the impact of e-Stroke on door-in-door-out time (DIDO), door-in to referral time (D2R) and 3-month modified Rankin Score (mRS) in this study.

Methods

The data was obtained from prospective thrombectomy registry between 1-Jan-2019 and 31-Mar-2021. The e-Stroke was implemented on 1-Mar-2020. The outcomes were compared between the period before (1-Jan-2019 to 28-Feb-2020) and after (1-Mar-2020 to 31-Mar-2021) implementation (Before-AI vs After-AI). No other changes to the pathway were made over this period. Welch’s t-test was used to compare time metrics and Fisher’s exact test for dichotomised mRS 0-2.

Results

Before-AI, 19 of 22 patients referred for MT were transferred. After-AI, 21 of 25 patients referred were transferred. The mean DIDO and D2R Before-AI vs After-AI were 141 vs 79 (p=0.001) and 71 vs 44 minutes (p=0.01) respectively. Dichotomized mRS 0-2 at 3 months was 16% vs 48% (p=0.04) before-AI vs after-AI. (Fig.1)

Figure1: 3-month mRS distribution before and after e-Stroke

Conclusions

Incorporating e-Stroke decision aid tool into our PSC hyperacute stroke pathway led to a significant reduction in door-in-door-out and door to referral times. A significantly higher proportion of patients gained functional independence at 3 months following the implementation of e-Stroke.

Background and Aims

Manual localization and quantification of silent brain infarcts is costly and challenging for clinicians, especially for studies where an accurate outcome is of enormous importance. The two most important challenges in the automated detection of infarcts, consist in the high class-imbalance and the presence of considerable white matter hyper-intensities, which are often miss-classified by automatic lesion segmentation methods. For that purpose we develop a deep learning framework which is capable of detecting these lesions.

Methods

The framework essentially consists of two modules. In the first step, a synergistic multi-scale network (the χ-Net) is applied, which fuses the complementary information contained in DWI and ADC. The network consists of two contracting paths and two up-sampling paths, whereby the information from the two down-sampling processes is fused. The resulting synergistic segmentation masks are further refined by an additional network block reducing the false positive rate by explicitly learning the difference between ischaemia and other hyper-intensity.

Results

Training and validation of the networks was carried out using data from patients who suffered stroke and admitted to our Radiology department. The proposed framework achieves a sensitivity of 0.9755 and a Dice coefficient of 0.8222.

Conclusions

The proposed χ-Net architecture with additional peeling module delivers promising results in stroke segmentation task. In order to be able to make more precise statistical statements about performance, we will carry out the experiments on a larger data set or on other medical imaging problems, where images which are derived from different modalities carry complementary information (e.g. CT and MR perfusion data).