Target-based deep learning network surveillance of non-contrast computed tomography for small infarct core of acute ischemic stroke

Abstract

Purpose: Rapid diagnosis of acute ischemic stroke (AIS) is critical to achieve positive outcomes and prognosis. This study aimed to construct a model to automatically identify the infarct core based on non-contrast-enhanced CT images, especially for small infarcts.

Methods: The baseline CT scans of AIS patients, who had DWI scans obtained within less than 2 h apart, were included in this retrospective study. A modified Target-based deep learning model of YOLOv5 was developed to detect infarctions on CT. Randomly selected CT images were used for testing and evaluated by neuroradiologists and the model, using the DWI as a reference standard. Intraclass correlation coefficient (ICC) and weighted kappa were calculated to assess the agreement. The paired chi-square test was used to compare the diagnostic efficacy of physician groups and automated models in subregions. p < 0.05 was considered statistically significant.

Results: Five hundred and eighty four AIS patients were enrolled in total, finally 275 cases were eligible. Modified YOLOv5 perform better with increased precision (0.82), recall (0.81) and mean average precision (0.79) than original YOLOv5. Model showed higher consistency to the DWI-ASPECTS scores (ICC = 0.669, κ = 0.447) than neuroradiologists (ICC = 0.452, κ = 0.247). The sensitivity (75.86% vs. 63.79%), specificity (98.87% vs. 95.02%), and accuracy (96.20% vs. 91.40%) were better than neuroradiologists. Automatic model had better diagnostic efficacy than physician diagnosis in the M6 region (p = 0.039).

Conclusion: The deep learning model was able to detect small infarct core on CT images more accurately. It provided the infarct portion and extent, which is valuable in assessing the severity of disease and guiding treatment procedures.

Keywords: acute ischemic stroke; non-contrast CT; small infarct core; target-based deep learning network; you only look once (YOLO).

Figure 1

The structure of the YOLOv5 model.

Figure 2

Parameter change curves for different loss functions during the training process. M-CIoU is in color blue, CIoU is in color pink and the DIoU is in gray.

Figure 3

Visualization of acute core infarct recognition on NCCT images using the deep learning model for case 1 and case 2. Case 1 was shown at the top and Case 2 at the bottom. Case 1 and 2 images were shown as NCCT (labeled as 1A and 2A), automatic model diagnosed images (labeled as 1B and 2B), DWI images (labeled as 1C and 2C) and ADC images(labeled as 1D and 2D).

Figure 4

Visualization of acute core infarct recognition on NCCT images using the deep learning model for case 3. Case 3 images were shown as NCCT (labeled as A), automatic model diagnosed images (labeled as B), DWI images (labeled as C) and ADC images (labeled as D).

Figure 5

The technical pipeline of the study.The study begins with the registration of CT and DWI images using the Dual Attention VoxelMorph Network. Radiologists then labeled the core infarcts on the CT images, which were subsequently used to train the Modified YOLOv5-based model. Finally, the trained model efficiently recognizes acute infarct cores and calculates the ASPECTS score.