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

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.

Simultaneous R2* and quantitative susceptibility mapping measurement enables differentiation of thrombus hematocrit and age: an in vitro study at 3 T.

BACKGROUND: The efficacy of acute ischemic stroke treatment is affected by thrombus composition and age, yet no diagnostic method capable of quantitative thrombus characterization currently exists. This in vitro study evaluates the use of R2* , quantitative susceptibility mapping (QSM), and proton density fat fraction (FF) maps derived from a single gradient echo (GRE) MRI acquisition for characterizing clot of various hematocrit, as well as added calcified and lipidic components, throughout aging. METHODS: Two thrombus phantoms containing porcine clots (10-60% hematocrit, one with added calcium or lard) were scanned serially throughout 6 days of aging. Three-dimensional multi-echo GRE imaging was used to generate R2* , QSM, and FF maps, from which mean values for all clots at every time point were obtained. Receiver operating characteristic analysis was used to derive thresholds differentiating acute from chronic clot, and measured R2* and QSM were tested for their ability to estimate clot hematocrit. RESULTS: R2* and QSM varied minimally over the first 6 hours of aging (acute), and QSM was found to linearly relate to clot hematocrit. Beyond 6 hours (chronic), R2* and QSM increased considerably over time and hematocrit could be estimated from the R2* /QSM ratio. R2* and QSM thresholds of 22 s-1 and 0.165 ppm differentiated acute from chronic clots with a sensitivity/specificity of 100%/100% and 85%/92%, respectively. QSM and FF maps definitively distinguished calcium and lipid, respectively, from clots of any hematocrit and age. CONCLUSIONS: R2* , QSM, and FF from a single multi-echo GRE scan discriminated hematocrit and age, and distinguished calcification and lipid withinin vitro clot.

Single multi-echo GRE acquisition with short and long echo spacing for simultaneous quantitative mapping of fat fraction, B0 inhomogeneity, and susceptibility.

Multi-echo gradient echo (mGRE) sequences have been widely adapted in clinical and scientific practice for different purposes to their capability of performing Dixon MRI, generating multi-contrast images and extracting multi-parametric maps. This work aims to extend mGRE-based techniques for imaging whole head, where further technical developments are required due to the co-existence of fat and large B0 inhomogeneity in regions such as the skull base and neck. Specifically, bipolar mGRE data were acquired with a single sequence that contains both a short echo-spacing (ΔTE) echo train to capture water-fat and B0 phase shifts (for proton density fat-fraction (FF) and B0 mapping) and a longer ΔTE echo train (and long echo times) to capture subtle susceptibility variations and R2* information. The mGRE images covering the whole head (spatial resolution 1.0 × 1.0 × 2.0 mm3) were acquired in 5 min. An automated processing pipeline was implemented to use the FF and B0 maps determined from the short-TE train to compensate for the effects of fat, remove the background phase for whole-head quantitative susceptibility mapping, and reduce the difficulty of spatial phase unwrapping of the long echo-time data. Data from healthy volunteers imaged on a 3 T scanner along with phantom validation are presented. Co-registered quantitative multi-parametric maps (FF, B0 inhomogeneity, R2*, local frequency shift and quantitative susceptibility) and multi-contrast images covering the whole head were successfully generated in processing times of several minutes.