How to increase field of view of CNN in stardist?

Hello,
I am currently trying to train a 3D model to segment some very elongated nuclei in a particularly crowded tissue. I manually annotated a couple of crops and I am now trying to play around with some stardist notebooks (like the ones on the Stardist GitHub page and from the ZeroCost colab page ) to understand what are the parameter that I need.

When I check the median size of my nuclei and the field of view (fov), I always get a message saying that the fov is too small. How do I increase it?
Example:

empirical anisotropy of labeled objects = (7.644736842105263, 1.6049723756906078, 1.0)
Config3D(anisotropy=(7.644736842105263, 1.6049723756906078, 1.0), axes=‘ZYXC’, backbone=‘resnet’, grid=(4, 4, 4), n_channel_in=1, n_channel_out=241, n_dim=3, n_rays=240, net_conv_after_resnet=128, net_input_shape=(None, None, None, 1), net_mask_shape=(None, None, None, 1), rays_json={‘name’: ‘Rays_GoldenSpiral’, ‘kwargs’: {‘n’: 240, ‘anisotropy’: (7.644736842105263, 1.6049723756906078, 1.0)}}, resnet_activation=‘relu’, resnet_batch_norm=False, resnet_kernel_init=‘he_normal’, resnet_kernel_size=(3, 3, 3), resnet_n_blocks=4, resnet_n_conv_per_block=3, resnet_n_filter_base=32, train_background_reg=0.0001, train_batch_size=1, train_checkpoint=‘weights_best.h5’, train_checkpoint_epoch=‘weights_now.h5’, train_checkpoint_last=‘weights_last.h5’, train_dist_loss=‘mae’, train_epochs=400, train_foreground_only=0.9, train_learning_rate=0.0003, train_loss_weights=(1, 0.2), train_n_val_patches=None, train_patch_size=(32, 256, 256), train_reduce_lr={‘factor’: 0.5, ‘patience’: 40, ‘min_delta’: 0}, train_steps_per_epoch=100, train_tensorboard=True, use_gpu=False)
Using default values: prob_thresh=0.5, nms_thresh=0.4.
Median object size: [ 19. 90.5 145.25]
FOV: [49 49 49]
WARNING: median object size larger than field of view of the neural network.

I read that one possibility is to scale down the stacks in all dimensions and increase the train patch size. I still get the same problem, but besides this, isn’t it a loss of information in the z axis?

Thank you!
Best,
Lucrezia

Please see this. In addition, you can also increase the receptive field by increasing the depth of the U-Net.

You can downscale the stacks differently for the spatial and axial dimensions, e.g. by factor 4 in X and Y and factor 2 in Z.

Best,
Uwe