Skip to main content
. Author manuscript; available in PMC: 2023 Jan 1.
Published in final edited form as: J Invest Dermatol. 2021 Jul 13;142(1):97–103. doi: 10.1016/j.jid.2021.06.015

Figure 1. Experimental workflow.

Figure 1.

a) Image stacks generation by RCM devices and biopsy validated consensus ground-truth generation by a panel of expert confocalists. b) CNN model used in this study. A ResNet34 pretrained backbone was used along with two extra layers of residual blocks to increase the receptive field of the last feature map. The numbers indicate the number of residual blocks in each of the layers. In addition to the final classifier, during training, a classification loss was backpropagated also from intermediate activation maps. Abbreviations: Avg Pool: average pooling; FC: fully connected layer; CE: Cross-entropy loss. Ground-truth abbreviations: N: normal skin; NB: not BCC; S: suspicious; B: BCC; BQ: bad quality image. Scale bar: a) 250 μm.