Integrated backscatter versus spectral parameters for in vivo estimation of human carotid plaque composition

Carotid plaque composition is a missing piece of information in the treatment of carotid stenosis. This study evaluates a spectral analysis-based approach versus an intensity only approach (comparable to an idealized grayscale) using both fundamental and harmonic bandwidths. The intensity approach utilizes the integrated backscatter (IB), while the spectral analysis approach uses the slope, intercept and mid-band fit from an estimate of the backscatter transfer function. Backscattered ultrasound RF data were acquired in vivo from 134 subjects prior to carotid endarterectomy. Serial histology slides of the surgically excised plaque were matched to grayscale images created from the RF data. 1.2 mm by 1.2 mm regions of interest (ROI) were selected in the RF data corresponding to homogenous regions within the histology, determined as calcified (Ca), fibrous or fibro-lipidic (F), and hemorrhagic and/or necrotic core (HNC). A balanced data set of 130 Ca, 120 F, and 125 HNC ROI's was randomly selected to train and test three random forest classifiers relying on 1) IB, 2) spectral linear fit parameters, or 3) both IB and spectral linear fit parameters as inputs. Color-coded maps for 30 randomly selected matched frames were produced from the classifiers and compared to the matched histology based on a blinded expert review. HNC accuracy and specificity were slightly better for the spectral linear fit based approach than IB (accuracy, 0.63 ± 0.05 vs 0.57 ± 0.06 and specificity, 0.77 ± 0.1 vs 67 ± 0.1), while sensitivity was the same for both (0.36 ± 0.07). The spectral linear fit parameter-based model provided the best representation of the plaque for 24 of 30 frames. These results support the understanding that spectral information can improve on the performance of intensity only based approaches for ultrasound-based tissue characterization of carotid plaque.