Clot Analog Attenuation in Non-contrast CT Predicts Histology: an Experimental Study Using Machine Learning.

Exact histological clot composition remains unknown. The purpose of this study was to identify the best imaging variables to be extrapolated on clot composition and clarify variability in the imaging of thrombi by non-contrast CT. Using a CT-phantom and covering a wide range of histologies, we analyzed 80 clot analogs with respect to X-ray attenuation at 24 and 48 h after production. The mean, maximum, and minimum HU values for the axial and coronal reconstructions were recorded. Each thrombus underwent a corresponding histological analysis, together with a laboratory analysis of water and iron contents. Decision trees, a type of supervised machine learning, were used to select the primary variable altering attenuation and the best parameter for predicting histology. The decision trees selected red blood cells (RBCs) for correlation with all attenuation parameters (p < 0.001). Conversely, maximum attenuation on axial CT offered the greatest accuracy for discriminating up to four groups of clot histology (p < 0.001). Similar RBC-rich thrombi displayed variable imaging associated with different iron (p = 0.023) and white blood cell contents (p = 0.019). Water content varied among the different histologies but did not in itself account for the differences in attenuation. Independent factors determining clot attenuation were the RBCs (β = 0.33, CI = 0.219-0.441, p < 0.001) followed by the iron content (β = 0.005, CI = 0.0002-0.009, p = 0.042). Our findings suggest that it is possible to extract more and valuable information from NCCT that can be extrapolated to provide insights into clot histological and chemical composition.