Unsupervised Deep Learning for Stroke Lesion Segmentation on Follow-up CT Based on Generative Adversarial Networks

Abstract

Background and purpose: Supervised deep learning is the state-of-the-art method for stroke lesion segmentation on NCCT. Supervised methods require manual lesion annotations for model development, while unsupervised deep learning methods such as generative adversarial networks do not. The aim of this study was to develop and evaluate a generative adversarial network to segment infarct and hemorrhagic stroke lesions on follow-up NCCT scans.

Materials and methods: Training data consisted of 820 patients with baseline and follow-up NCCT from 3 Dutch acute ischemic stroke trials. A generative adversarial network was optimized to transform a follow-up scan with a lesion to a generated baseline scan without a lesion by generating a difference map that was subtracted from the follow-up scan. The generated difference map was used to automatically extract lesion segmentations. Segmentation of primary hemorrhagic lesions, hemorrhagic transformation of ischemic stroke, and 24-hour and 1-week follow-up infarct lesions were evaluated relative to expert annotations with the Dice similarity coefficient, Bland-Altman analysis, and intraclass correlation coefficient.

Results: The median Dice similarity coefficient was 0.31 (interquartile range, 0.08-0.59) and 0.59 (interquartile range, 0.29-0.74) for the 24-hour and 1-week infarct lesions, respectively. A much lower Dice similarity coefficient was measured for hemorrhagic transformation (median, 0.02; interquartile range, 0-0.14) and primary hemorrhage lesions (median, 0.08; interquartile range, 0.01-0.35). Predicted lesion volume and the intraclass correlation coefficient were good for the 24-hour (bias, 3 mL; limits of agreement, -64-59 mL; intraclass correlation coefficient, 0.83; 95% CI, 0.78-0.88) and excellent for the 1-week (bias, -4 m; limits of agreement,-66-58 mL; intraclass correlation coefficient, 0.90; 95% CI, 0.83-0.93) follow-up infarct lesions.

Conclusions: An unsupervised generative adversarial network can be used to obtain automated infarct lesion segmentations with a moderate Dice similarity coefficient and good volumetric correspondence.

FIG 1.

The FU2BL-GAN global architecture (asterisk). The follow-up (FU) NCCT with lesion is clipped between Hounsfield unit ranges of 0−100 and 100−1000 and normalized to (−1) (double asterisks). The original BL NCCT is only clipped between 0 and 100 HU and normalized to (−1). The FU NCCT with a lesion is passed through the generator network to compute a difference map. This difference map is subtracted from the input FU NCCT to construct a generated BL NCCT. Original BL and generated BL are optimized on the basis of the absolute voxelwise difference (L1-loss) and the binary cross-entropy loss (adversarial-loss) of the discriminator networks classification (original or generated BL).

FIG 2.

Patients included in the training, validation, and test sets. The training data consisted of a BL and at least 1 follow-up (FU) NCCT. FU of <8 hours: FU NCCT acquired within 8 hours; FU 24H: FU NCCT acquired 8–72 hours; FU 1W: FU NCCT acquired 72  hours to 2 weeks after endovascular treatment or randomization. Validation and test sets were constructed with data from the studies by Konduri et al and Hssayeni et al. 8H indicates 8 hours.

FIG 3.

Dice similarity coefficients of test sets: 24-hour follow-up after AIS (24H infarct), 1-week follow-up after AIS (1W infarct), HT, and PrH. A, The results of all the test set data. B, Only results from lesions that are >10 mL. Each shade of color represents the results based on the supervised nnUnet approach, the FU2BL-GAN approach trained with L1+adv, and the generator trained with L1-loss only (L1) respectively. The Asterisk indicates P < .05; double asterisks, P < .001; triple asterisks, P < 1e-10; NS, nonsignificant difference.

FIG 4.

Bland-Altman plots of predicted lesion size for the FU2BL-GAN. A, 24H infarct follow-up. B, 1W infarct follow-up. C, HT. D, PrH.

FIG 5.

Visual results of the FU2BL-GAN. The first column contains the input NCCT with lesion used as input for the generator model to generate a difference map (column 2). The difference map is subtracted from the input NCCT (column 1) to obtain a generated BL scan (column 3). The negative (blue) and positive (red) values of the difference map correspond to the deviation of the difference map from zero. A higher deviation from zero implies a higher attenuation adjustment of the follow-up NCCT to generate the BL NCCT without a lesion. Column 4 contains the ground truth lesion annotations. Arrows show false-positive hemorrhage (rows 3 and 4), false-negative infarct (row 5, upper arrow), false-positive infarct (row 5, lower arrow), and false-negative hemorrhage segmentation (arrow, row 6).