Predicting final ischemic stroke lesions from initial diffusion-weighted images using a deep neural network

Abstract

Background: For prognosis of stroke, measurement of the diffusion-perfusion mismatch is a common practice for estimating tissue at risk of infarction in the absence of timely reperfusion. However, perfusion-weighted imaging (PWI) adds time and expense to the acute stroke imaging workup. We explored whether a deep convolutional neural network (DCNN) model trained with diffusion-weighted imaging obtained at admission could predict final infarct volume and location in acute stroke patients.

Methods: In 445 patients, we trained and validated an attention-gated (AG) DCNN to predict final infarcts as delineated on follow-up studies obtained 3 to 7 days after stroke. The input channels consisted of MR diffusion-weighted imaging (DWI), apparent diffusion coefficients (ADC) maps, and thresholded ADC maps with values less than 620 × 10^-6 mm²/s, while the output was a voxel-by-voxel probability map of tissue infarction. We evaluated performance of the model using the area under the receiver-operator characteristic curve (AUC), the Dice similarity coefficient (DSC), absolute lesion volume error, and the concordance correlation coefficient (ρ_c) of the predicted and true infarct volumes.

Results: The model obtained a median AUC of 0.91 (IQR: 0.84-0.96). After thresholding at an infarction probability of 0.5, the median sensitivity and specificity were 0.60 (IQR: 0.16-0.84) and 0.97 (IQR: 0.93-0.99), respectively, while the median DSC and absolute volume error were 0.50 (IQR: 0.17-0.66) and 27 ml (IQR: 7-60 ml), respectively. The model's predicted lesion volumes showed high correlation with ground truth volumes (ρ_c = 0.73, p < 0.01).

Conclusion: An AG-DCNN using diffusion information alone upon admission was able to predict infarct volumes at 3-7 days after stroke onset with comparable accuracy to models that consider both DWI and PWI. This may enable treatment decisions to be made with shorter stroke imaging protocols.

Keywords: Acute ischemic stroke; DWI; Deep learning; Lesion segmentation; MRI; PWI.

Fig. 1

Flow diagram of the study.

Fig. 2

The block diagram of the attention-gated U-net, as well as the network’s input and output.

Fig. 3

Examples of CNN model prediction compared with ADC thresholding in different reperfusion groups including major (A), minimal (B), partial (C), and unknown (D). The Dice similarity coefficients (DSC) shown below the images were calculated compared to the ground truth in all slices. Green area shows true positives, blue area false negatives, and red area false positives. (For interpretation of the references to colour in this figure legend, the reader is referred to the web version of this article.)

Fig. 4

The correlation of lesion volumes from the model prediction and ground truth manually delineated infarct size at 3–7 days after stroke, plotted using the cube root of the lesion sizes (ρ_c = 0.73, p < 0.01). In each subset, the solid lines represent the best linear fit function and the colored areas represent the 95 % confidence interval.

Fig. 5

Bland-Altman plots for patients in different reperfusion groups including major (A), minimal (B), partial (C), and unknown (D). The X-axis represents the mean volume, and the Y-axis represents the volume difference between the predicted and ground truth lesions. The solid line represents the bias, and the dashed lines represent the upper and lower 95% limits of agreement. Note that the error increases for larger lesions.

Fig. 6

The correlation between the DSC and cubic-root baseline lesion volumes (ρ = 0.64, p < 0.01). The best fit to an exponential function is indicated as a black curve.