all-zero output [deepmedic]

[YilinLiu97]

Hi, I retrained the deepmedic model with my dataset. The training looks fine, but the segmented output is either all-zero or like this:

My config file:

[T1]
path_to_search = /home/yilinliu/Dataset/Training
filename_contains =
filename_not_contains =
spatial_window_size = (57,57,57)
pixdim = (1.0, 1.0, 1.0)
axcodes=(A, R, S)
interp_order = 3

[parcellation]
path_to_search = /home/yilinliu/Dataset/FullStructureLabels
filename_contains = _RIGHT_all_labels_8bit_path
filename_not_contains =
spatial_window_size = (9,9,9)
pixdim = (1.0, 1.0, 1.0)
axcodes=(A, R, S)
interp_order = 0

%system configuration sections
[SYSTEM]
cuda_devices = 0
num_threads = 2
num_gpus = 1
model_dir = ./models/deepmedic

[NETWORK]
name = deepmedic
activation_function = prelu
batch_size = 10
decay = 0
reg_type = L2

% volume level preprocessing
volume_padding_size = 12
% histogram normalisation
histogram_ref_file = ./example_volumes/monomodal_parcellation/standardisation_models.txt
norm_type = percentile
cutoff = (0.01, 0.99)
normalisation = False
whitening = True
normalise_foreground_only=True
foreground_type = otsu_plus
multimod_foreground_type = and

queue_length = 25
window_sampling = uniform

[TRAINING]
sample_per_volume = 1000
rotation_angle = (-10.0, 10.0)
scaling_percentage = (-10.0, 10.0)
lr = 0.01
loss_type = CrossEntropy
starting_iter = 0
save_every_n = 3
max_iter = 500
max_checkpoints = 1000

[INFERENCE]
border = (24,24,24)
#inference_iter = 294
save_seg_dir = ./segmentation_output/
output_interp_order = 0
spatial_window_size = (105,105,105)

I've tried using different iterations for the inference, but the segmentation results show no improvement.

Any help is appreciated!!

[wyli]

Hi @YilinLiu97 The config looks ok to me. I'd suggest that you start with simple tasks, e.g. training with a few volumes (or even synthetic examples) and making inferences on the training cases. In that way we can narrow down the issue -- if there is any.

[YilinLiu97]

Hi @wyli , I've tried using just 1 subject as the input and it just outputs zeros (though the loss has been decreased to ~0.008.) Also I'm using the training sample for inference as the 'exclude_fraction_for_validation' function doesn't seem to be implemented in the 0.2.1 version or it's simply not included in the config file.

[MancaZerovnikMekuc]

I have the same problem as @wyli . The loss on validation as well as on the training set is getting really low but the result after inference is mostly 0. Is this because of shape constraints?

[fepegar]

Probably. Have you read the U-Net shape tutorial?

[MancaZerovnikMekuc]

Thank you. Yes I have read it. I have solved my problem with the help of #76 . As suggested there I had changed my loss function from Dice to generalised Dice (GDSC) since my data is imabalanced.