Deep learning prediction of stroke thrombus red blood cell content from multiparametric MRI

Abstract

Background and purpose: Thrombus red blood cell (RBC) content has been shown to be a significant factor influencing the efficacy of acute ischemic stroke treatment. In this study, our objective was to evaluate the ability of convolutional neural networks (CNNs) to predict ischemic stroke thrombus RBC content using multiparametric MR images.

Materials and methods: Retrieved stroke thrombi were scanned ex vivo using a three-dimensional multi-echo gradient echo sequence and histologically analyzed. 188 thrombus R₂*, quantitative susceptibility mapping and late-echo GRE magnitude image slices were used to train and test a 3-layer CNN through cross-validation. Data augmentation techniques involving input equalization and random image transformation were employed to improve network performance. The network was assessed for its ability to quantitatively predict RBC content and to classify thrombi into RBC-rich and RBC-poor groups.

Results: The CNN predicted thrombus RBC content with an accuracy of 62% (95% CI 48-76%) when trained on the original dataset and improved to 72% (95% CI 60-84%) on the augmented dataset. The network classified thrombi as RBC-rich or poor with an accuracy of 71% (95% CI 58-84%) and an area under the curve of 0.72 (95% CI 0.57-0.87) when trained on the original dataset and improved to 80% (95% CI 69-91%) and 0.84 (95% CI 0.73-0.95), respectively, on the augmented dataset.

Conclusions: The CNN was able to accurately predict thrombus RBC content using multiparametric MR images, and could provide a means to guide treatment strategy in acute ischemic stroke.

Keywords: Ischemic stroke; MRI; RBC; deep learning; machine learning; thrombus imaging.

Figure 1.

Representative training set thrombus RBC content distribution (A) before and (B) after input sampling equalization.

Figure 2.

An example thrombus (A) R₂*, (B) QSM, (C) GRE magnitude image slice, shown inside its scanning vial, along with the (D) FIESTA-C image used for segmentation. The resulting normalized, segmented thrombus 3-channel image slice used as CNN input is shown (E) before and (F) after random image transformation. R₂*, QSM and GRE magnitude pixel values represent the red, green and blue image color channels, respectively.

Figure 3.

Linear regression plots of the CNN predicted thrombus RBC content against the histological value for the networks trained on the (A) original and (B) augmented datasets. Plotted are the median predictions from all MR slices available for each thrombus. These same predictions replotted with only thrombi with histological RBC content between 20–45% are shown in (C) and (D), respectively.

Figure 4.

ROC curves for the identification of RBC-poor thrombi based on the quantitative RBC content predictions of the network trained on the original and augmented datasets.