Question about Bilater_voting in connect_loss.py

[JaronTu]

Hi, thanks for your brilliant work! I'd like to ask about "pred_mask, _ = torch.max(vote_out, dim=2)" in function Bilater_voting in connect_loss.py. In my opinion, this transforms the connectivity map to prediction map, but what is the logic of this operation?

[Zyun-Y]

Good question. Yes, it transforms the connectivity map into the prediction map.

The logic is that as long as one direction gets connected from the current pixel (since we are taking max), we assume it is part of the connected region (i.e., salient/positive).

You might see a difference in our original paper about this part where we used an 8-channel average to get the prediction map - the logic behind that one was that we wanted an average high probability of the connection from all its 8-direction neighboring pixels. Both work well but in our later studies [1][2], we found doing the max one gives even better results globally.

A similar idea can be found in the pooling methods: max-pooling and average-pooling - max doing better since it simply gets the strongest signal.

[1]Z. Yang, S. Soltanian-Zadeh, K. K. Chu, H. Zhang, L. Moussa, A. E. Watts, N. J. Shaheen, A. Wax, and S. Farsiu, “Connectivity-based Deep Learning Approach for Segmentation of the Epithelium in In Vivo Human Esophageal OCT Images,” Biomed. Opt. Express, vol. 12, no. 10, pp. 6326-6340, 2021.
[2] Z. Yang, S. Farsiu, "Directional Connectivity-based Segmentation of Medical Images", in CVPR 2023

[JaronTu]

I get it! Thanks for your detailed explanation!