Experimentations

Experimentations

The experiments use the Bullitt_Isotrope_light_mult32 dataset.
The experiments run on a single GTX 1080 8Gb GPU.

Unet model parameters and hyperparameter experimentation

This experimentation aims to highlighting the advantages and disadvantages of the different parameters and hyperparameters on a 3D unet model.

Model specifications

• DFT (Default : unet_exp_1 with default parameters)
• DFZ (Double Filter Size : filters_mult = 2)
• QFZ (Quadruple Filter Size : filters_mult = 4)
• OFZ (Octuple Filter Size : filters_mult = 8)
• 5KS (5 Kernel Size : change kernel size to (5 x 5 x 5))
• SPD (SPatial Dropout : use SpatialDropout3D instead of Dropout)

Results

• train loss and valid loss : on patchs
• model prediction dice coef : on a full image prediction

Patchs (64 x 64 x 64) (unet_exp_1)

patch_size_x = 64
patch_size_y = 64
patch_size_z = 64
batch_size = 2
steps_per_epoch = 500
epochs = 100

Model	time for 100 epochs	best model train loss	best model valid loss	best model prediction dice coef
DFT	10132 s	-0.64114	-0.65603	0.63938
DFZ	19329 s	-0.66005	-0.66816	0.65519
QFZ	53524 s	-0.68017	-0.68382	0.66635
5KS	25211 s	-0.64559	-0.65732	0.63030
SPD	9868 s	-0.63552	-0.65994	0.62955

Patchs (64 x 64 x 64) (unet_exp_2)

patch_size_x = 64
patch_size_y = 64
patch_size_z = 64
batch_size = 2
steps_per_epoch = 500
epochs = 100

Model	time for 100 epochs	best model train loss	best model valid loss	best model prediction dice coef
DFT	10870 s	-0.64565	-0.65816	0.63917
DFZ	21789 s	-0.66256	-0.66992	0.65613

Patchs (96 x 96 x 96) (unet_exp_1)

patch_size_x = 96
patch_size_y = 96
patch_size_z = 96
batch_size = 1
steps_per_epoch = 500
epochs = 100

Model	time for 100 epochs	best model train loss	best model valid loss
DFT	18207 s	-0.65676	-0.66486

Patchs (16 x 16 x 16) (unet_exp_1)

patch_size_x = 16
patch_size_y = 16
patch_size_z = 16
batch_size = 2
steps_per_epoch = 1500
epochs = 100

Model	time for 100 epochs	best model train loss	best model valid loss	best model prediction dice coef
DFT	4841 s	-0.19408	-0.26721	0.54538
DFZ	5223 s	-0.24959	-0.27506	0.54672
QFZ	11909 s	-0.24956	-0.28085	0.56449

Patchs (96 x 96 x 96) (unet_exp_4)

The predictions are made with different patchs size : 16 x 16 x 16, 32 x 32 x 32, 64 x 64 x 64 ; and with the full image.

patch_size_x = 96
patch_size_y = 96
patch_size_z = 96
batch_size = 1
steps_per_epoch = 500
epochs = 100

Model	time for 100 epochs	best model train loss	best model valid loss	patchs 16	patchs 32	patchs 64	full image
DFT	54054 s	-0.68892	-0.68825	0.63767	0.66143	0.66382	0.66443