Fig. 2 | Scientific Reports

Fig. 2

From: Evaluating the generalizability of an automated coronary artery calcium segmentation and scoring algorithm using multi-vendor dataset

Fig. 2

Visualization of the per-lesion segmentation performance of RICAU-Nets on images containing noise (a), noise and CAC (b), and artificial implants with CAC (c). The lesions are zoomed-in in the images. The Grad-CAM23 attention maps were generated with respect to the bone category, as both noise and artificial implants were categorized within this category. The per-lesion Dice scores for each image are as follows: background: 99.97%, bone: 98.76% for (a); background: 99.96%, bone: 98.14%, RCA: 90.24% for (b); background: 99.98%, bone: 99.73%, LAD: 100% for (c). The red heatmap areas indicating noise and artificial implants demonstrated that RICAU-Nets effectively distinguished between these elements and calcium lesions.

Back to article page