Skip to main content
. Author manuscript; available in PMC: 2023 Mar 27.
Published in final edited form as: Med Image Anal. 2022 Dec 14;84:102726. doi: 10.1016/j.media.2022.102726

Fig. 3.

Fig. 3.

The framework of the proposed method for deep feature interpretation. Given any deep feature f on the neural network, it is up-sampled to the size of the input image. Then the resized feature map is binarized by using the proposed prototype segmentation (ProtoSeg) with the initial mask B. The segmentation ability of the feature f is measured by the SA score between the segmentation map Sf and the ground-truth G. Our proposed ProtoSeg can be used in any deep feature or input image to measure their segmentation abilities and reveal the transition of the segmentation ability from the input image to the output segmentation.