Figure 7.

Network multi-slice (a) training and (b) testing framework. Training with three slices is shown as an example. (a) Multi-slice inputs reconstructed from sparse projection data is processed with the masking procedure described in Fig. 8. The generated segmentation maps are shared with multi-slice targets reconstructed from full projection data. Patches are extracted as training samples only when they contain more than 50% foreground pixels based on the generated masks, termed tissue of interest (TOI) oriented. (b) Five consecutive testing slices are used to reconstruct the central slice, indicated by the yellow bounding box. Three sets of multi-slice inputs, where the target slice has different slice context, are independently processed by the same trained network. Only the target slices are retained and aggregated to obtain the final reconstruction of the target slice.