Two-stage CNNs for computerized BI-RADS categorization in breast ultrasound images

Table 3 Comparisons of one-stage models and two-stage models with the grading accuracy of each category

	Category 3	Category 4A	Category 4B	Category 4C	Category 5
Refined ROI-CNN + G-CNN	0.998 ± 0.0040	0.940 ± 0.0110	0.734 ± 0.0662	0.922 ± 0.0376	0.876 ± 0.1234
Refined ROI-CNN + VGG	0.990 ± 0.0047	0.920 ± 0.0194	0.673 ± 0.0692.	0.908 ± 0.0493	0.841 ± 0.1319
Refined ROI-CNN + ResNet	0.991 ± 0.0052	0.927 ± 0.0155	0.688 ± 0.0665	0.920 ± 0.0388	0.858 ± 0.1423
ROI-CNN + G-CNN	0.955 ± 0.0076	0.897 ± 0.0112	0.679 ± 0.0678	0.906 ± 0.0434	0.837 ± 0.1232
ROI-CNN + VGG	0.947 ± 0.0145	0.864 ± 0.0132	0.667 ± 0.0726	0.865 ± 0.0463	0.818 ± 0.1281
ROI-CNN + ResNet	0.954 ± 0.0136	0.878 ± 0.0124	0.669 ± 0.0664	0.899 ± 0.0425	0.835 ± 0.1338
One-stage G-CNN	0.797 ± 0.0190	0.552 ± 0.0312	0.496 ± 0.0876	0.715 ± 0.0674	0.559 ± 0.1257
One-stage VGG	0.723 ± 0.0244	0.533 ± 0.0471	0.460 ± 0.0888	0.644 ± 0.0535	0.436 ± 0.1328
One-stage ResNet	0.755 ± 0.0268	0.550 ± 0.0302	0.472 ± 0.0799	0.692 ± 0.0585	0.508 ± 0.1402

One-stage models were consisted of experimental test cases “One-stage G-CNN”, “One-stage VGG” and “One-stage ResNet”. Two-stage models involved six experiment test cases, including “Refined ROI-CNN + G-CNN”, “Refined ROI-CNN + VGG”, “Refined ROI-CNN + ResNet”, “ROI-CNN + G-CNN”, “ROI-CNN + VGG”, and “ROI-CNN + ResNet”
The significance for the italic values is to illustrate the method with the best performance according to each evaluated metric

ISSN: 1475-925X