Deep learning biomarker of chronometric and biological ischemic stroke lesion age from unenhanced CT

Abstract

Estimating progression of acute ischemic brain lesions - or biological lesion age - holds huge practical importance for hyperacute stroke management. The current best method for determining lesion age from non-contrast computerised tomography (NCCT), measures Relative Intensity (RI), termed Net Water Uptake (NWU). We optimised lesion age estimation from NCCT using a convolutional neural network - radiomics (CNN-R) model trained upon chronometric lesion age (Onset Time to Scan: OTS), while validating against chronometric and biological lesion age in external datasets (N = 1945). Coefficients of determination (R²) for OTS prediction, using CNN-R, and RI models were 0.58 and 0.32 respectively; while CNN-R estimated OTS showed stronger associations with ischemic core:penumbra ratio, than RI and chronometric, OTS (ρ² = 0.37, 0.19, 0.11); and with early lesion expansion (regression coefficients >2x for CNN-R versus others) (all comparisons: p < 0.05). Concluding, deep-learning analytics of NCCT lesions is approximately twice as accurate as NWU for estimating chronometric and biological lesion ages.

Fig. 1. Methods overview.

a Two ischemic lesion segmentation CNN models were trained using expert-annotated 3D NCCT images: Acute (AC: Onset Time to Scan, OTS: <6 h); and Subacute (SA: OTS: 6–48 h). b Validation of lesion segmentations occurred in two ways: (i) experts provided with Test NCCT images and clinical syndromes, judged whether the 1^st-ranked SubAcute model segmented component overlay the relevant ischemic area (as for Image #001); unless no relevant SubAcute output existed, when the Acute model was judged instead (#002). If the 1^st-ranked component was incorrect, e.g. lesion unrelated to presentation, or artefact, then the 2nd component was interrogated (#003), and so for the 3rd, but no further components; (ii) a random subsample of segmented, adjudicated components were directly compared with each of two experts’ manual annotations. c Lesion age (OTS) model was derived from a CNN generated from multiple patches of expert-annotated ischemic regions (using unregistered NCCT images). The CNN-OTS estimate was combined with radiomic features from the same ischemic region and its flipped control image (using spatially realigned and normalized images). A Relative Intensity (RI) model, equivalent to lesion Net Water Uptake (NWU), was also developed from normalized difference in mean NCCT attenuation within ischemic and flipped masks. The CNN model, supplemented with radiomic features (CNN-R), and RI model, were validated using segmented ischemic lesions from the Test cohort, as per b. d Model Training, Calibration and Test were performed in three approximately equal number, independent datasets, with Training images derived from different hospitals to those used for Calibration and Test; and with the former using expert-drawn lesions, and the latter using automatically-segmented, adjudicated, lesions. Calibration and Test sets were alternated (i, ii) to generate OTS estimates of all 1547 subjects. Smaller Calibration sets (0–500), and correspondingly larger Test sets, were also evaluated.

Fig. 2. Lesion segmentation validation.

a Flow chart of Test cohort, showing numbers of expert-adjudicated ischemic lesions identified by either segmentation model, as a function of whether the lesion was identified by the 1^st, 2^nd or 3^rd probabilistically-ranked model output component. Lower ranked components were considered segmentation failures. Successfully segmented lesions were then used for validation of i) lesion segmentation spatial similarity (see b); and ii) lesion age model compared to chronometric age (OTS), and two measures of lesion biological age. b Lesion segmentation fidelity was evaluated in random subsets of 70 acute (OTS < 6 hours: left graphs) and 70 subacute (OTS 6–48 h: right graphs) ischemic lesions identified by one of the top three segmentation model outputs, compared to two expert annotations. Top two graphs show Dice Similarity Coefficients (DSC) of model versus experts, as well as inter-expert comparison, as a function of expert annotation volume, with log function lines of best fit shown. Lower graphs show scatter plots of lesion volume comparing expert annotations versus model (Auto) outputs. Horizontal bars reflect the range of experts. Circles denote thin slice images; triangles denote thick slice images.

Fig. 3. Chronometric OTS validation.

Scatterplots of lesion age estimates in Test cohort as a function of actual chronometric OTS. First two models shown use Relative Intensity (Net Water Uptake) alone, either as a linear or a Gaussian Process Regression (GPR) function. Third model uses the CNN model supplemented with radiomic features (CNN-R) developed within the Training cohort. Axes are logarithmic.

Fig. 4. Core:penumbra ratio estimation.

a Scatterplots of lesion age estimates from paired NCCT-CT Perfusion (CTP) Test subset, comparing against CTP ischemic core : penumbra ratio mismatch. Age estimates originate from the optimal Relative Intensity (GPR) and CNN-R models. Correlation coefficient of plot using CNN-R was greater than that using RI model: Z = 2.62, p < 0.01. The grey zone represents cases where penumbra : core ratio is <1.2 (a threshold used to select patients for revascularization). Axes are logarithmic. b Scatterplots of a subset of cases where chronometric OTS was also available, showing superiority of CNN-R estimated age relative to both RI model and chronometric OTS (Z > 2.7; p < 0.01).

Fig. 5. Infarct growth prediction.

a Scatterplots of ischemic lesion volumes in Validation subset with paired NCCTs at two time points (white and black circles respectively), representing predominantly scans on admission, and 24–48 h follow up. Axes are logarithmic. b Comparison of regression coefficients for lesion age (at first scan) as a predictor within a multiregression model of percent ischemic lesion expansion. Lesion age value is taken as either chronometric OTS, or as the estimates from optimized Relative Intensity (RI) or CNN-R models. Only the CNN-R age estimate coefficient compared to the other two age estimates was significant (Z > 2.14; p < 0.05). c Standardized residuals of a baseline model of percent ischemic lesion expansion plotted as a function of lesion age estimates (at first scan). Negative-y axis values indicate that estimated infarct growth is less than that estimated by a model using baseline predictors of: lesion volume at t1; t2-t1 interval, and stroke severity. The three graphs use chronometric OTS, RI, and CNN-R estimated ages respectively. Correlation coefficient of plot using CNN-R estimated age was significantly greater than plots using the other two age estimates (Z ≥ 3.24; p < 0.01).