ICCV2021-CDNet: Centripetal Direction Network for Nuclear Instance Segmentation
[ICCV2021]
The code includes training and inference procedures for CDNet.
Tips: There is a result written mistake (U-Net) in Table 4 in the original paper. The correct result is：
| Method Name | Dice | AJI |
|---|---|---|
| U-Net | 0.8184 | 0.5910 |
| Mask-RCNN | 0.7600 | 0.5460 |
| DCAN | 0.7920 | 0.5250 |
| Micro-Net | 0.7970 | 0.5600 |
| DIST | 0.7890 | 0.5590 |
| CIA-Net | 0.8180 | 0.6200 |
| U-Net | 0.8027 | 0.6039 |
| Hover-Net | 0.8260 | 0.6180 |
| BRP-Net | - | 0.6422 |
| PFF-Net | 0.8091 | 0.6107 |
| Our CDNet | 0.8316 | 0.6331 |
cd CDNet/
python train.py
cd CDNet/
python test.py
The size of the training set is 300, and the size of the test set is 1000, what is the reason for this？
Hi guys, you have implemented many models, but which one of them is/are actually the CDNet? As default training model you have specified a model, whose name is not CDNet so I got a bit confused :)
你好，请问monuseg数据集你们是怎么处理的，可以提供一份吗
您好，复现您的代码时，使用的是(MoNuSeg, CPM17) 两个dataset，但是一直报错，
Traceback (most recent call last):
File "/media/kemove/ericgu/CDNet-main/train.py", line 528, in
main()
File "/media/kemove/ericgu/CDNet-main/train.py", line 293, in main
dsets[x] = DataFolder(dir_list, post_fix, num_channels, data_transforms[x])
File "/media/kemove/ericgu/CDNet-main/data_folder.py", line 94, in init
raise(RuntimeError('Found 0 image pairs in given directories.'))
RuntimeError: Found 0 image pairs in given directories.
为什么要设置三个train_300文件夹，还需要对(MoNuSeg, CPM17)再进行什么处理吗？
我看您的文章还对图像进行了一些label
希望您能详细指点，谢谢！
Good job. Could you provide the format of the dataset? I don't understand the meaning of weight_map in the dataset
您好，可以再次提供下您的百度网盘数据集吗，那个链接访问人数超出，不能下载了
Hi, I am trying to reproduce the results using the default settings, but the performance is too low.
Are there any special hyper-parameters to reproduce the results on the MoNuSeg dataset?
Also, I would like to know if there is any plan to release any pretrained model.
Thanks.
======>>>True, 0_UNet2RevA1_vgg16[None][adam]_sche[None]_3c_input256over40bs8_e300_MSE_addDir
================================= seed = 2022 =================================
2022-04-15 07:57 ***** Training starts *****
2022-04-15 07:57 save directory: ./experiments/MoNuSeg_oridata/0_UNet2RevA1_vgg16[None][adam]_sche[None]_3c_input256over40bs8_e300_MSE_addDir
2022-04-15 07:57
------------------------- Options -------------------------
dataset: MoNuSeg_oridata
isTrain: True
all_img_test: 1
momentum: 0.95
*************** model ***************
multi_class: True
in_c: 3
out_c: 3
direction: 1
n_layers: 6
growth_rate: 24
drop_rate: 0.1
compress_ratio: 0.5
is_hybrid: True
layer_type: basic
mean_std: mean_std
add_weightMap: 1
dice: 1
boundary_loss: 0
mseloss: 1
modelName: UNet2RevA1_vgg16
backbone: None
pretrained: 1
LossName: CE1_Dice1
exp_filename: 0_UNet2RevA1_vgg16[None][adam]_sche[None]_3c_input256over40bs8_e300_MSE_addDir
direction_classes: 9
*************** train ***************
branch: 5
num_epochs: 300
input_size: 256
batch_size: 8
val_overlap: 40
seed: 2022
early_stop: 0
scheduler: None
step: 5
lr: 0.001
lr_decay: 0.995
weight_decay: 0.0001
log_interval: 15
workers: 8
gpu: [0]
alpha: 0.0
optimizer: adam
validation: 0
checkpoint_freq: 100
start_epoch: 0
checkpoint:
trans_train: _noRRe_isRCo_isHF_noRA_isRE_noRRo_isRCr_isCAu_isLE_noNorm
data_dir: ./data/MoNuSeg_oridata
save_dir: ./experiments/MoNuSeg_oridata/0_UNet2RevA1_vgg16[None][adam]_sche[None]_3c_input256over40bs8_e300_MSE_addDir
weight_map_dir: ./data/MoNuSeg_oridata/weight_maps
img_dir: ./data/MoNuSeg_oridata/images
label_dir: ./data/MoNuSeg_oridata/labels
transform_string: _noRRe_isRCo_isHF_noRA_isRE_noRRo_isRCr_isCAu_isLE_noNorm
transform_str: _noRRe_isRCo_isHF_noRA_isRE_noRRo_isRCr_isCAu_isLE_noNorm
*************** transform ***************
train:
random_color: 1
horizontal_flip: True
vertical_flip: True
random_elastic: [6, 15]
random_chooseAug: 1
random_crop: 256
label_encoding: [3, 2, 1]
to_tensor: 1
val:
label_encoding: [3, 2, 1]
to_tensor: 1
-------------------------- End --------------------------
CrossEntropyLoss()
opt.train[trans_train] = _noRRe_isRCo_isHF_noRA_isRE_noRRo_isRCr_isCAu_isLE_noNorm
upsample_blocks[0] in: 512 out: 256
upsample_blocks[1] in: 256 out: 128
upsample_blocks[2] in: 128 out: 64
upsample_blocks[3] in: 64 out: 32
upsample_blocks[4] in: 32 out: 16
Scheduler None not available
====== Compose(_random_color_horizontal_flip_vertical_flip_random_elastic_random_chooseAug_random_crop_label_encoding_to_tensor), ====== each epoch work time is [0.0000 min].
====== Compose(_label_encoding_to_tensor), ====== each epoch work time is [0.0000 min].
dir_list = ['./data/MoNuSeg_oridata/images/train_300', './data/MoNuSeg_oridata/weight_maps/train_300', './data/MoNuSeg_oridata/labels/train_300']
post_fix = ['weight.png', 'label.png']
2022-04-15 07:58 =======================No validation=======================
2022-04-15 07:58 Epoch: [1/300]
2022-04-15 07:58 Iteration: [0/24] Loss 8.5217 loss_direction_CE 2.7267 loss_direction_dice 1.0622 loss_mse 0.8598 Loss_CE 1.6661 Loss_var -1.0000 Pixel_Accu 0.8939
pixel_IoU 0.0159 pixel_Recall 0.0587 pixel_Precision 0.0228 pixel_F1 0.0310
2022-04-15 07:58 Iteration: [15/24] Loss 6.3783 loss_direction_CE 1.7026 loss_direction_dice 1.0091 loss_mse 0.6421 Loss_CE 1.1705 Loss_var -1.0000 Pixel_Accu 0.9321
pixel_IoU 0.0132 pixel_Recall 0.0461 pixel_Precision 0.0403 pixel_F1 0.0252
2022-04-15 07:58 => Train Avg: Loss 5.9872 loss_direction_CE 1.5172 loss_direction_dice 1.0017 loss_mse 0.6094 Loss_CE 1.0626 Loss_var -1.0000 Pixel_Accu 0.9472
pixel_IoU 0.0096 pixel_Recall 0.0316 pixel_Precision 0.0530 pixel_F1 0.0184
======= train() ======== each epoch work time is [18.0887 s].
======= validate() ======== each epoch work time is [0.0001 s].
.
.
.
2022-04-15 09:30 Epoch: [300/300]
2022-04-15 09:30 Iteration: [0/24] Loss 3.6795 loss_direction_CE 0.7567 loss_direction_dice 0.9106 loss_mse 0.3836 Loss_CE 0.7377 Loss_var -1.0000 Pixel_Accu 0.9829
pixel_IoU 0.3683 pixel_Recall 0.4962 pixel_Precision 0.5779 pixel_F1 0.5307
2022-04-15 09:30 Iteration: [15/24] Loss 3.5008 loss_direction_CE 0.7322 loss_direction_dice 0.8279 loss_mse 0.3279 Loss_CE 0.6730 Loss_var -1.0000 Pixel_Accu 0.9788
pixel_IoU 0.3732 pixel_Recall 0.5125 pixel_Precision 0.5754 pixel_F1 0.5371
2022-04-15 09:30 => Train Avg: Loss 3.5091 loss_direction_CE 0.7307 loss_direction_dice 0.8309 loss_mse 0.3267 Loss_CE 0.6702 Loss_var -1.0000 Pixel_Accu 0.9788
pixel_IoU 0.3695 pixel_Recall 0.5104 pixel_Precision 0.5715 pixel_F1 0.5328
epoch = 299, cp_flag = 1
2022-04-15 09:30 This model is: [UNet2RevA1_vgg16]. add_weightMap = 1. The time spent is [92.482 min].
------------------------- Options -------------------------
dataset: MoNuSeg_oridata
isTrain: False
all_img_test: 1
momentum: 0.95
*************** model ***************
multi_class: True
in_c: 3
out_c: 3
direction: 1
n_layers: 6
growth_rate: 24
drop_rate: 0.1
compress_ratio: 0.5
is_hybrid: True
layer_type: basic
mean_std: mean_std
add_weightMap: 1
dice: 1
boundary_loss: 0
mseloss: 1
modelName: UNet2RevA1_vgg16
backbone: None
pretrained: 1
LossName: CE1_Dice1
exp_filename: 0_UNet2RevA1_vgg16[None][adam]_sche[None]_3c_input256over40bs8_e300_MSE_addDir
direction_classes: 9
transform_string: ['random_color', 'random_chooseAug', 'horizontal_flip', 'random_elastic', 'random_crop', 'label_encoding', 'to_tensor']
transform_str: _noRRe_isRCo_isHF_noRA_isRE_noRRo_isRCr_isCAu_isLE_noNorm
*************** transform ***************
test:
to_tensor: 1
normalize: [[0.68861804 0.46102882 0.61138992],
[0.19204499 0.20979484 0.1658672 ]]
*************** post ***************
postproc: 0
min_area: 20
radius: 2
*************** test ***************
filename: test1
epoch: best
gpu: [0]
branch: 5
groundtruth: 0
img_dir: ./data/MoNuSeg_oridata/images/test1
label_dir: ./data/MoNuSeg_oridata/labels/test1
annotation_dir: ./data/MoNuSeg_oridata/Annotations
weight_map_dir: ./data/MoNuSeg_oridata/weight_maps
tta: True
save_flag: True
patch_size: 256
overlap: 40
savefilename: br5_test1_gt0_post0_best_minarea20_ra2
save_dir: ./experiments/MoNuSeg_oridata/0_UNet2RevA1_vgg16[None][adam]_sche[None]_3c_input256over40bs8_e300_MSE_addDir/br5_test1_gt0_AIT1_post0_best_minarea20_ra2
model_path: ./experiments/MoNuSeg_oridata/0_UNet2RevA1_vgg16[None][adam]_sche[None]_3c_input256over40bs8_e300_MSE_addDir/checkpoints/checkpoint_best.pth.tar
model_branch_path: ./experiments/MoNuSeg_oridata/0_UNet2RevA1_vgg16[None][adam]_sche[None]_3c_input256over40bs8_e300_MSE_addDir/checkpoints/checkpointBranch_best.pth.tar
model_branch_path2: ./experiments/MoNuSeg_oridata/0_UNet2RevA1_vgg16[None][adam]_sche[None]_3c_input256over40bs8_e300_MSE_addDir/checkpoints/checkpointBranch2_best.pth.tar
save_testfilename: br5_test1_gt0_AIT1_post0_best_minarea20_ra2
-------------------------- End --------------------------
====== Compose(_to_tensor_normalize), ====== each epoch work time is [0.0000 min].
MoNuSeg_oridata_logExl.csv have been exist.
upsample_blocks[0] in: 512 out: 256
upsample_blocks[1] in: 256 out: 128
upsample_blocks[2] in: 128 out: 64
upsample_blocks[3] in: 64 out: 32
upsample_blocks[4] in: 32 out: 16
=> loading trained model
=> loaded model at epoch 245
=> Test begins:
img_names changes => : ['TCGA-21-5784-01Z-00-DX1.png', 'TCGA-21-5786-01Z-00-DX1.png', 'TCGA-B0-5698-01Z-00-DX1.png', 'TCGA-B0-5710-01Z-00-DX1.png', 'TCGA-CH-5767-01Z-00-DX1.png', 'TCGA-E2-A14V-01Z-00-DX1.png', 'TCGA-E2-A1B5-01Z-00-DX1.png', 'TCGA-G9-6362-01Z-00-DX1.png']
=> Processing image TCGA-21-5784-01Z-00-DX1.png
label_img_instance.len = 399
Computing output probability maps...
===> Processed all 0 images, ===> The every test time spent is [2.00 s]
Computing metrics...
Ns == 22, Ng == 398
Count = 14.0
[ana_FP = 0.0001, ana_FN = 0.0388, ana_less = 0.9553, ana_more = 0.0058]
image TCGA-21-5784-01Z-00-DX1.png, the pixel_iou = 0.0096, pixel_recall = 0.0097, pixel_precision = 0.6215, pixel_F1 = 0.0191
(0.035175879396976084, 0.6363636363607439, 0.06666666666665079, 0.42350635464529324, 0.294697861901088, 15.708433531593744, 0.00964156489459278), result_AJI = 0.00964156489459278, result_Dice = 0.01909898567933733, result_Dice2 = 0, dq_value = 0.009523809523809525, sq_value = 0.6711873476907836, pq_value = 0.006392260454197939
Saving image results...
pred_labeled color number = 23
=> Processing image TCGA-21-5786-01Z-00-DX1.png
label_img_instance.len = 441
Computing output probability maps...
===> Processed all 1 images, ===> The every test time spent is [0.15 s]
Computing metrics...
Ns == 137, Ng == 440
Count = 104.0
[ana_FP = 0.0030, ana_FN = 0.2583, ana_less = 0.7284, ana_more = 0.0103]
image TCGA-21-5786-01Z-00-DX1.png, the pixel_iou = 0.0591, pixel_recall = 0.0592, pixel_precision = 0.9772, pixel_F1 = 0.1116
(0.23636363636358265, 0.7591240875906868, 0.3604852686307867, 0.29449895800721115, 0.18495710548718708, 20.72356024523777, 0.048276896763836584), result_AJI = 0.04956267347685077, result_Dice = 0.11157164366230242, result_Dice2 = 0, dq_value = 0.01386481802426343, sq_value = 0.618584259587985, pq_value = 0.008576558191861143
Saving image results...
pred_labeled color number = 138
=> Processing image TCGA-B0-5698-01Z-00-DX1.png
label_img_instance.len = 358
Computing output probability maps...
===> Processed all 2 images, ===> The every test time spent is [0.15 s]
Computing metrics...
Ns == 7, Ng == 357
Count = 4.0
[ana_FP = 0.0007, ana_FN = 0.0126, ana_less = 0.9852, ana_more = 0.0016]
image TCGA-B0-5698-01Z-00-DX1.png, the pixel_iou = 0.0016, pixel_recall = 0.0016, pixel_precision = 0.4144, pixel_F1 = 0.0032
(0.011204481792713949, 0.5714285714204081, 0.02197802197801594, 0.21051631359475842, 0.12370657595857719, 23.66638498771461, 0.001588895125016798), result_AJI = 0.001588895125016798, result_Dice = 0.0031727490844803636, result_Dice2 = 0, dq_value = 0.0, sq_value = 0.0, pq_value = 0.0
Saving image results...
pred_labeled color number = 8
=> Processing image TCGA-B0-5710-01Z-00-DX1.png
label_img_instance.len = 360
Computing output probability maps...
===> Processed all 3 images, ===> The every test time spent is [0.15 s]
Computing metrics...
Ns == 64, Ng == 359
Count = 46.0
[ana_FP = 0.0039, ana_FN = 0.0796, ana_less = 0.9014, ana_more = 0.0151]
image TCGA-B0-5710-01Z-00-DX1.png, the pixel_iou = 0.0428, pixel_recall = 0.0436, pixel_precision = 0.7100, pixel_F1 = 0.0821
(0.12813370473534036, 0.7187499999988769, 0.21749408983446394, 0.5330615112466484, 0.3909037328718037, 9.797606772605755, 0.0421813327620162), result_AJI = 0.04276215508078669, result_Dice = 0.08207090673329327, result_Dice2 = 0, dq_value = 0.05673758865248227, sq_value = 0.6768533029673144, pq_value = 0.03840302428183344
Saving image results...
pred_labeled color number = 65
=> Processing image TCGA-CH-5767-01Z-00-DX1.png
label_img_instance.len = 295
Computing output probability maps...
===> Processed all 4 images, ===> The every test time spent is [0.16 s]
Computing metrics...
Ns == 303, Ng == 295
Count = 192.0
[ana_FP = 0.0181, ana_FN = 0.5555, ana_less = 0.3515, ana_more = 0.0750]
image TCGA-CH-5767-01Z-00-DX1.png, the pixel_iou = 0.2439, pixel_recall = 0.2543, pixel_precision = 0.8565, pixel_F1 = 0.3922
(0.650847457626898, 0.6336633663364245, 0.6421404682273173, 0.4510931922447862, 0.3255492017367811, 14.887090364552593, 0.20432282594444756), result_AJI = 0.20663469037955884, result_Dice = 0.3922006220839813, result_Dice2 = 0, dq_value = 0.18394648829431437, sq_value = 0.630962549086058, pq_value = 0.11606334514960932
Saving image results...
pred_labeled color number = 304
=> Processing image TCGA-E2-A14V-01Z-00-DX1.png
label_img_instance.len = 379
Computing output probability maps...
===> Processed all 5 images, ===> The every test time spent is [0.14 s]
Computing metrics...
Ns == 320, Ng == 378
Count = 169.0
[ana_FP = 0.1326, ana_FN = 0.4191, ana_less = 0.3098, ana_more = 0.1384]
image TCGA-E2-A14V-01Z-00-DX1.png, the pixel_iou = 0.4439, pixel_recall = 0.4814, pixel_precision = 0.8507, pixel_F1 = 0.6149
(0.44708994708982885, 0.528124999999835, 0.4842406876790137, 0.6254490587294964, 0.49236794632726577, 13.625374921253902, 0.33986179928309423), result_AJI = 0.3461042482675505, result_Dice = 0.6149068322981367, result_Dice2 = 0, dq_value = 0.2521489971346705, sq_value = 0.6842801543181587, pq_value = 0.17254055467048127
Saving image results...
pred_labeled color number = 321
=> Processing image TCGA-E2-A1B5-01Z-00-DX1.png
label_img_instance.len = 330
Computing output probability maps...
===> Processed all 6 images, ===> The every test time spent is [0.15 s]
Computing metrics...
Ns == 353, Ng == 329
Count = 271.0
[ana_FP = 0.1934, ana_FN = 0.5260, ana_less = 0.0864, ana_more = 0.1942]
image TCGA-E2-A1B5-01Z-00-DX1.png, the pixel_iou = 0.5173, pixel_recall = 0.6078, pixel_precision = 0.7764, pixel_F1 = 0.6818
(0.8237082066866798, 0.7677053824360431, 0.7947214076245168, 0.6852851280441974, 0.5602336428794454, 8.406038574890323, 0.4476577529206426), result_AJI = 0.4649849842381133, result_Dice = 0.6818385827811073, result_Dice2 = 0, dq_value = 0.5219941348973607, sq_value = 0.7106749201016873, pq_value = 0.3709681401117312
Saving image results...
pred_labeled color number = 354
=> Processing image TCGA-G9-6362-01Z-00-DX1.png
label_img_instance.len = 473
Computing output probability maps...
===> Processed all 7 images, ===> The every test time spent is [0.16 s]
Computing metrics...
Ns == 481, Ng == 472
Count = 333.0
[ana_FP = 0.0343, ana_FN = 0.5654, ana_less = 0.3106, ana_more = 0.0898]
image TCGA-G9-6362-01Z-00-DX1.png, the pixel_iou = 0.3747, pixel_recall = 0.3960, pixel_precision = 0.8745, pixel_F1 = 0.5452
(0.7055084745761218, 0.6923076923075484, 0.6988457502622561, 0.6036795700795706, 0.46884170695346195, 11.052712051937652, 0.33817976219031626), result_AJI = 0.3402095310317029, result_Dice = 0.5451589031063631, result_Dice2 = 0, dq_value = 0.3378803777544596, sq_value = 0.666468963341415, pq_value = 0.22518678509542042
Saving image results...
pred_labeled color number = 482
scale_size_notmatch = []
===> Processed all 8 images, ===> The time spent is [2.01 min]
Average of all images:
pixel_accu: 0.8475 pixel_IoU 0.2116 pixel_Recall 0.2317 pixel_Precision 0.7602 pixel_F1 0.3063
recall: 0.3798 precision: 0.6634 F1: 0.4108 dice: 0.4784 iou: 0.3552 haus: 14.7334
AJI_sklearn: 0.1323
AJI: 0.1790 hover_AJI: 0.1827 hover_Dice: 0.3063
result_Dice2: 0.0000, dq_value: 0.1720, sq_value: 0.5824, pq_value: 0.1173 analysis_FP: 0.0482, analysis_FN: 0.3069, analysis_pred_less: 0.5786, analysis_pred_more: 0.0663
Wow! The illustration presented on the front page of the paper is so great! I tried to find the relevant code implementation in the repository, but unfortunately I couldn't find it.
And I tried to draw the subgraph (i) in Figure 1 myself, but i couldn't project it on the X-Y plane. Therefore, to seek some help, if the relevant code implementation or the simplified implementation process could be told, will be very grateful!
你好，我昨天下载了那两个数据集。
请问我该如何在项目路径中放置这两个数据集？我尝试按readme中的来但有问题。
预处理需要提前做吗？还是直接将原始数据按路径放置代码就能跑起来了？
感谢！
Hi, Can you please reshare datasets links as data isn't available on the links shared earlier.
Hello, may I ask how the data preparation weight_maps is obtained? Is there any relevant code?
Hello, I'm trying to run your excellent CDNet on a customized dataset. It looks like I need to specify 'data_dir' and 'weight_map_dir' in options.py. Could you explain how you organize the data, especially what is in the two directories? And could you point me to the code you preprocess the data? Thank you!
React
A declarative, efficient, and flexible JavaScript library for building user interfaces.
🖖 Vue.js is a progressive, incrementally-adoptable JavaScript framework for building UI on the web.
Typescript
TypeScript is a superset of JavaScript that compiles to clean JavaScript output.
An Open Source Machine Learning Framework for Everyone
Django
The Web framework for perfectionists with deadlines.
Laravel
A PHP framework for web artisans
Bring data to life with SVG, Canvas and HTML. 📊📈🎉
JavaScript (JS) is a lightweight interpreted programming language with first-class functions.
Some thing interesting about web. New door for the world.
A server is a program made to process requests and deliver data to clients.
Machine learning is a way of modeling and interpreting data that allows a piece of software to respond intelligently.
Some thing interesting about visualization, use data art
Some thing interesting about game, make everyone happy.
Facebook
We are working to build community through open source technology. NB: members must have two-factor auth.
Microsoft
Open source projects and samples from Microsoft.
Google
Google ❤️ Open Source for everyone.
Alibaba
Alibaba Open Source for everyone
D3
Data-Driven Documents codes.
Tencent
China tencent open source team.