Meeting minutes of the weekly meeting with Tom on the 26 Oct 2016:
What I am doing:
==========
Replicating the MICCAI results to send them to Max for the TMI journal paper.
RSG meeting:
========
We will organise a meeting with the Leuven guys when they arrive on the 28th November, Valentina will be involved on this as well.
TODO:
====
– Test dilated kernels for higher accuracy of the FCN rather than pooling layers. We would have to assess the impact of this improvement on the
inference speed.
– Fine-tune the network with a Dice LOSS layer.
– Two ideas for registration:
1) Iterate over CNN, warping, CNN, warping, CNN, warping… And put a regulariser.
2) Use the uncertainty for the aforementioned regularisation of the iterative process. Uncertainty is given by using dropout at inference time.
Dropout is used to generate different samples of the output and measure the changes in the output in those samples, if the outputs are
very different for a particular pixel it means that there is a high uncertainty.
Transfer viva:
========
Write the outline of the thesis. Deep learning in image guided fetal interventions.
1) Segmentation of surgical tools in real-time: visual servoing, mosaicking and skills assessment.
2) Registration in real-time: propagation of segmentations, registering slices of the fetal MR while scanning.
3) 3D reconstruction: endoscopic video, photoacoustic.