We want to create a custom pipeline where the input is a set of 3D images and the output shape nd array is the same shape(3D) but with N number of channels. This is for a heatmap based approach to solve a landmark detection code. We are using Clara 3.1 (moving to 4.0 is not an option for us). Can anyone help us on what kind of pipeline should be used? We are thinking to reuse the segmentation pipeline and make changes to the loss and data pipeline. The loss function, data loading and the architecture will be customised.
Reference paper: Integrating spatial configuration into heatmap regression based CNNs for landmark localization, Christian Payer et al, Med. Image Analysis’ 19