Ahmad Mustafa, Ghassan AlRegib, Reza Rastegar

3D CNNs can exploit the full extent of spatial information in seismic volumes to predict faults. They require large quantities of training data, but this issue has been mitigated by training such networks with large amounts of synthetic training data to apply afterwards on real datasets. Because of domain shift, pre-trained networks may fail to perform as expected, emphasizing the need to finetune the network with labels obtained on target data of interest. The mismatch in dimensions of network input and interpreter annotations (mostly 2D slices) poses a problem in this respect. In addition, there is a high degree of uncertainty attached to such labels, both on a coarse level as well as on a more fine-grained basis. On an image level, the interpreter may only annotate structures in a small, localized portion of the seismic volume. Furthermore, owing to uncertainty regarding the exact delineation of the endpoints of picked faults, interpreters may only label certain segments on the complete fault line. We propose a method whereby we demonstrate a procedure to finetune the pretrained 3D CNNs with sparse 2D labels on target datasets, resulting in the adaption of their weights to better pick faults in the new domains. Secondly, we devise means to incorporate interpretation uncertainty on labels produced for finetuning to generate better, more reliable fault estimations. We validate our findings on various real datasets and demonstrate improved network performance over the pretrained CNN alone. Additionally, we show that incorporating label uncertainty while finetuning leads to better interpretation performance compared to uncertainty-agnostic finetuning.