HyLFM-Net-stat

statically trained HyLFM-Net

This network was trained on light-field and corresponding light-sheet data acquired with a HyLFM microscope—a hybrid light-field light-sheet microscope. For a description of the implementation and information about the microscope please refer to "Deep learning-enhanced light-field imaging with continuous validation." at https://doi.org/10.1038/s41592-021-01136-0. For implementation details and how to train such networks please additionally refer to https://github.com/kreshuklab/hylfm-net.

This network is not expected to transfer to data acquired with a different microscope. Even if your light-field data recorded medaka heart nuclei with a 20x objective represented with $n_{num}=19$, this model might not yield faithful reconstructions.

With a HyLFM microscope continuous validation is enabled by the acquisition of groundtruth-like light-sheet planes. This network should only be used in such a hybrid optical setup to ensure its reconstruction bias toward Medaka heart nuclei is justified and the implicitly learned noise and scattering distributions are applicable.

training command:

hylfm/train.py heart_static_fish2_f4 --batch_multiplier=2 --batch_size=1 --c00_2d=768 --c00_3d=32 --c01_2d=768 --c02_2d=512 --c03_2d=256 --c10_2d=512 --c10_3d=8 --c11_2d=256 --c11_3d=8 --c12_2d=0 --cin_3d=64 --crit_beta=1 --crit_decay_weight_by=0.8949745643842519 --crit_decay_weight_every_unit=epoch --crit_decay_weight_every_value=1 --crit_ms_ssim_weight=0.01 --crit_threshold=2.394979173266996 --crit_weight=0.007110228990791768 --criterion=WeightedSmoothL1_MS_SSIM --init_fn=xavier_uniform --kernel2d=3 --kernel3d=3 --last_kernel2d=5 --lr_sched_factor=0.5 --lr_sched_patience=8 --lr_sched_thres=0.0001 --lr_sched_thres_mode=abs --lr_scheduler=ReduceLROnPlateau --max_epochs=200 --nnum=19 --opt_lr=0.0002663980703785583 --optimizer=Adam --patience=15 --up0_2d=384 --up0_3d=16 --validate_every_unit=epoch --validate_every_value=2 --z_out=49