SK-Project Brain Hemorrhage Segmentation

Information

Measures

Accuracy
DICE Coefficient
Jaccard Coefficient (IoU)
Slice Level Accuracy (For Segmentation 2D)
If model found Hemorrhage Pixel(at least 1 pixels), then get accuracy scores

Data Description

SK Dataset
Kaggle Dataset (RSNA Datasets) \ Kaggle RSNA Intracranial Hemorrhage Detection (https://www.kaggle.com/c/rsna-intracranial-hemorrhage-detection)

Logger

Train Logger : epoch, loss, IoU, Dice, Slice-Level-Accuracy
Test Logger : epoch, loss, IoU, Dice, Slice-Level-Accuracy
Classifications : epoch, loss, accuracy

Getting Started

Requirements

Python3 (3.6.8)
PyTorch (1.2)
torchvision (0.4)
NumPy
pandas
matplotlib
pytorch-efficientnet (Only Use Classificaitions)
tensorboard (Use Segmentation-2D)

Classifications - Train Examples

python3 main.py --loss-function bce --exp Classification/Classification --optim-function radam --momentum 0.9 --initial-lr 0.0001 --lr-schedule 75 100 --weight-decay 0.0001 --batch-size 24 --tenosrboardwriter Classifications/Unet-Encoder-Classification --model unet --aug False --smooth False

python main.py \
--batch_size 4 * n (n is Number of GPUs) \
--scheduler 75 100 \
--initial-lr 0.0001 \
--weight_decay 0.0001 \
--batch-size 24 \
--tenosrboardwriter Classifications/Unet-Encoder-Classification
--momentum 0.9 \
--model Unet \
--data sk-datasets \
--exp Classification/Classification \
--smooth False \
--aug False

Args	Options	Description
trn-root	[str]	dataset locations.
tst-root	[str]	dataset locations.
model	efficientnet, resnet	model architecture : efficientnet (defalut : b1 model), Resnet (defalut : 50), default : efficient net
batch_size	[int]	number of samples per batch. default : 8
epochs	[int]	number of epochs for training. default : 50
learning_rate	[float]	learning rate. defalut : 0.1
exp	[str]	./test/
number	[int]	a number of efficient net. default : b1
momentum	[int]	momentum of optimizers. default : 0.9
tenosrboardwriter	[str]	save path of tensor board
kaggle	[str]	Kaggle Datasets. If True use Kaggle Datasets, default : False

Segmentation - 2D Train Examples

python3 main.py --loss-function bce --exp Segmentation/Unet --optim-function radam --momentum 0.9 --initial-lr 0.0001 --lr-schedule 75 100 --weight-decay 0.0001 --batch-size 24 --tenosrboardwriter Segmentation/Unet-Encoder --model unet --aug False --smooth False --coordconv [none]

python main.py \
--batch_size 4 * n (n is Number of GPUs) \
--scheduler 75 100 \
--lr 0.0001 \
--optim-function radam \
--weight_decay 0.0001 \
--momentum 0.9 \
--model Unet \
--data sk-datasets \
--exp ./MTL/Segmentation/Unet/ \
--tenosrboardwriter Segmentation/Unet-Encoder \
--aug False \
--smooth False \
--coordconv [none]

Args	Options	Description
trn-root	[str]	dataset locations.
tst-root	[str]	dataset locations.
model	unet, unet coordconv, unet scse, unet multiinput	model architecture : unet base models, default : unet
batch_size	[int]	number of samples per batch. default : 8
epochs	[int]	number of epochs for training. default : 200
scheduler	[int]	100 170 200 epoch decay 0.1 defalut : 100 170 200
learning_rate	[float]	learning rate. defalut : 0.1
exp	[str]	./test/
momentum	[int]	momentum of optimizers. default : 0.9
tenosrboardwriter	[str]	save path of tensor board
coorconv	[list]	the number of coordconv layers. default : [9]

Segmentation - 3D Train Examples

python main.py --trn-root /data1/JM/sk_project/data2th_trainvalid_3d_patches_48_48_48_st_16_bg_0.1_nonzero_0.1 --val-root /data1/JM/sk_project/data2th_test_3d_patches_48_48_48_st_16_bg_1_nonzero_0.1 --input-size 48 48 48 --lr-schedule 20 30 --lr 0.01 --weight-decay 0.0005 --gamma 0.1 --f-maps 32 64 128 256 --stride-test 16 --exp jm_test1 --batch-size 128 --loss-function dice

python main.py \
--trn-root /data1/JM/sk_project/data2th_trainvalid_3d_patches_48_48_48_st_16_bg_0.1_nonzero_0.1 \
--val-root /data1/JM/sk_project/data2th_test_3d_patches_48_48_48_st_16_bg_1_nonzero_0.1 \
--input-size 48 48 48 \
--lr-schedule 20 30 \
--lr 0.01 \
--weight-decay 0.0005 \
--gamma 0.1 \
--f-maps 32 64 128 256 \
--stride-test 16 \
--exp exp_sk-project-3d \
--batch-size 128 \
--loss-function dice

Args	Options	Description
trn-root	[str]	dataset locations.
tst-root	[str]	dataset locations.
model	[str]	model architecture. default : unet
f-maps	[int]	feature map size of encoder layer. default : [32, 64, 128, 256]
conv-layer-order	[str]	order of layers. 'cr' -> conv + ReLU ,'crg' -> conv + ReLU + groupnorm , 'cbr' -> conv + BN + Relu. default : cbr
batch-size	[int]	number of samples per batch. default : 8
input-size	[int]	size of input patch. default : [48,48,48].
epochs	[int]	number of epochs for training. default : 200
lr-schedule	[int]	epoch decay 0.1. defalut : [20,30,35]
weight-decay	[float]	weight-decay. defalut : 0.0005
loss-function	[str]	bce ,dice, weight_bce. defalut : bce
learning_rate	[float]	learning rate. defalut : 0.1
exp	[str]	save folder name.

Reference

[1] U-Net: Convolutional Networks for Biomedical Image Segmentation (Ronneberrger et al.)
[2] 3D U-Net: Learning Dense Volumetric Segmentation from Sparse Annotation (Abdulkadir et al.)
[3] EfficientNet: Rethinking Model Scaling for Convolutional Neural Networks (Tan et al.)
[4] A Novel Focal Tversky loss function with improved Attention U-Net for lesion segmentation (Abraham and Khan)
[5] Concurrent Spatial and Channel Squeeze & Excitation in Fully Convolutional Network (Roy et al.)
[6] An intriguring falling of convolutional neural networks and the coordconv solution (Liu et al.) \

Geongyu / SK-Project