Table 3.

Overall accuracies of canine sarcoma spectra classification by the three CNN architectures.

Protocol	variant_Lecun	variant_LeNet	variant_VGG9
Scenario A	0.92 ± 0.01 (04%a 02%b)	0.95 ± 0.01 (08%a 03%b)	0.94 ± 0.01 (04%a 01%b)
Scenario B	0.95 ± 0.02 (08%a 05%b 03%c)	0.99 ± 0.00 (12%a 07%b 04%c)	0.96 ± 0.00 (06%a 03%b 02%c)

The best result for each task (accuracy ± standard variation over 10 independent iterations) is indicated in boldface.

^aThe improvement is expressed as a percentage relative to learning from scratch.

^bThe improvement is expressed as a percentage relative to transfer learning.

^cThe improvement is expressed as a percentage relative to Scenario A.