Automated carotid IMT measurement and its validation in low contrast ultrasound database of 885 patient Indian population epidemiological study: results of AtheroEdge™ Software

Abstract

Aim: The aim of this paper was to demonstrate the usage of an automated computer-based IMT measurement system called - CALEX 3.0 (a class of patented AtheroEdge™ software) on a low contrast and low resolution image database acquired during an epidemiological study from India. The image contrast was very low with pixel density of 12.7 pixels/mm. Further, to demonstrate the accuracy and reproducibility of the AtheroEdge™ software system we compared it with the manual tracings of a vascular surgeon--considered as a gold standard.

Methods: We automatically measured the IMT value of 885 common carotid arteries in longitudinal B-Mode images. CALEX 3.0 consisted of a stage for the automatic recognition of the carotid artery and an IMT measurement modulus made of a fuzzy K-means classifier. Performance was assessed by measuring the system accuracy and reproducibility against manual tracings by experts.

Results: CALEX 3.0 processed all the 885 images of the dataset (100% success). The average automated obtained IMT measurement by CALEX 3.0 was 0.407±0.083 mm compared with 0.429 ± 0.052 mm for the manual tracings, which led to an IMT bias of 0.022±0.081mm. The IMT measurement accuracy (0.022 mm) was comparable to that obtained on high-resolution images and the reproducibility (0.081 mm) was very low and suitable to clinical application. The Figure-of-Merit defined as the percent agreement between the computer-estimated IMT and manually measured IMT for CALEX 3.0 was 94.7%.

Conclusion: CALEX 3.0 had a 100% success in processing low contrast/low-resolution images. CALEX 3.0 is the first technique, which has led to high accuracy and reproducibility on low-resolution images acquired during an epidemiological study. We propose CALEX 3.0 as a generalized framework for IMT measurement on large datasets.

Figure 1

Samples of high-resolution and low-resolution images. A) High-resolution image where the intima and adventitia layers are neatly defined, noise is low, and pixel density is high (about 20 pixels/mm); B) image acquired by a medium-end scanner without compound imaging, where LI is hypoechoic; C) images acquired by a low-end equipment without harmonic and compound imaging, where LI is almost invisible and pixel resolution is very low (about 12 pixles/mm); D) example of low-resolution image with high level of image noise.

Figure 2

Samples of low-resolution images extracted from the dataset. Panels A, C, E show the carotid. Panels B, D, and F show the zoomed portion of the corresponding dashed rectangle on the left. The white arrows indicate the challenges of these images: interrupted intima representation (panel B), low contrast between intima, media, and adventitia (panel D), and well represented adventitia, but hypoechic intima (panel F).

Figure 3

CALEX 3.0 automated cropping. A) Original image. The dashed lines delimit the region containing the ultrasound data. Outside the dashed lines, the vertical and horizontal gradients are null. The white arrows on the right indicate the vertical scale for the measurement of the conversion factor. B) Cropped image.

Figure 4

Automated carotid identification (Stage-I) by CALEX 3.0. A) Original cropped image; B) line segments; C) line segments corresponding to the near and far adventitia layers obtained through validation and classification; D) final profile of the far adventitia (AD_F); E) determination of a Guidance Zone in which segmentation is performed (white dashed rectangle); F) extracted Guidance Zone of the distal wall.

Figure 5

Samples of CALEX 3.0 automated segmentation. The image is zoomed in the Guidance Zone.

Figure 6

Lumen region detection. A) Original cropped image; B) lumen region detection. The pixels possibly belonging to the lumen are mapped to white.

Figure 7

Samples of CALEX 3.0 automated segmentation (white lines) compared to manual segmentations (white dashed lines). The left column reports the lumen-intima profiles, the right the media-adventitia.

Figure 8

Correlation plot between CALEX 3.0 IMT values (vertical axis) and GT IMT values (horizontal axis).

Figure 9

Bland-Altmann plot of the CALEX 3.0 IMT measurements compared to ground truth.

Figure 10

Histogram distribution of the IMT measurement bias. It can be noticed that the average IMT bias is very low and the standard deviation of the histogram is lower than 0.1 mm.