Fig. 6: Schematic outlining the protocol optimisation dual-network strategy.
From: Dual deep learning approach for non-invasive renal tumour subtyping with VERDICT-MRI
We first remove the non-DW images from the image volumes, resulting in 27 measurements per patient. Our feature scoring network takes in the 27 measurements (27 input nodes), has hidden layers with 64 nodes and outputs 27 scores for the measurements. From these, the 12 measurements with the highest scores are selected. These are then used to reconstruct the target initial 27 measurements in the predictor network, and the loss is computed as the MSE between the output and the target. We evaluate on our full patient dataset, choosing b = 70, 150, 1000, 2000 as our final protocol. This gives a reduction in acquisition time of 32 min.
