CancerDetection

https://www.kaggle.com/c/histopathologic-cancer-detection

Dependencies

For install dependencies. Run

pip3 install -r requirements.txt

How to start

Install all dependencies and create configuration files.

If you want use our selections use this command:

git lfs pull

We have configuration files in research/data directory here

In this directory you can find README with description of configs.

For evaluation script use this command:

python3 research/evaluate/evaluation.py -s <config>

For training script use this command:

python3 research/train/trainer.py -s <config>

Testing

Models are presented here

Solo models

Number of expetiment	Augmentation	Network architecture	Additional params	Accuracy
1	A	ResNet18	Epoch 18	0.9028
2	A	ResNet18	Epoch 1	0.8897
3	A	MobileNetV2	Epoch 1	0.8892
4	A	MobileNetV2	Epoch 6	0.8850
5	A	MobileNet	Epoch 7	0.8839
6	A	ResNet50	Epoch 7	0.8948
7	A	ResNet50	Epoch 17	0.8949
8	B	ResNet18	Epoch 2	0.9016
9	B	ResNet18	Epoch 3	0.9044
10	B	ResNet18	Epoch 6	0.9072
11	B	ResNet18	Epoch 9	0.9080
12	B	ResNet18	Epoch 12	0.8985
13	B	ResNet18	Epoch 22	0.8878
14	B	DenseNet169	Epoch 3	0.8999
15	B	DenseNet169	Epoch 5	0.9084
16	B	DenseNet169	Epoch 9	0.9042
17	B	MeNet456	Epoch 13	0.8975
18	B	MeNet456	Epoch 18	0.8947
19	B	MeNet456	Epoch 20	0.9047
20	B	MeNet456	Epoch 21	0.9023
21	B	MeNet456	Epoch 29	0.9103
22	B	MeNet456	Epoch 31	0.9033
23	C	MeNet456*	Epoch 3 + Output 1x1 + Dropout	0.9510
24	C	Dense121*	Epoch 20(Last Layer) + Epoch 18 + Output 1x1 + Dropout	0.9745
25	C	Dense121*	Epoch 20(Last Layer) + Epoch 16 + Output 1x1 + Dropout	0.9742
26	C	Dense169*	Epoch 23 + Output 1x1 + Dropout	0.9710
27	C	Dense169*	Epoch 13 + Output 1x1 + Dropout	0.9695
28	C	ResNet18*	Epoch 17 + Output 1x1 + Dropout	0.9684
29	C	ResNet18*	Epoch 18 + Output 1x1 + Dropout	0.9685
30	C	Dense121*	Epoch 20 + Input 96x96 + Output 1x1 + Dropout	0.9655
31	C	Dense121*	Epoch 18 + Input 96x96 + Output 1x1 + Dropout	0.9632
32	C	Dense121*	Epoch 14 + Input 96x96 + Output 1x1 + Dropout	0.9665

• We use these layers instead fully-connected layer (layer with output 1000 ImageNet classes) for modeles marked *

    num_ftrs = net.fc.in_features
    out_ftrs = int(net.fc.out_features / 4)
    net.fc = nn.Sequential(
        nn.Sigmoid(),
        nn.Dropout(0.5),
        nn.Linear(num_ftrs, out_ftrs, bias=True),
        nn.SELU(),
        nn.Dropout(0.7),
        nn.Linear(in_features=out_ftrs, out_features=1, bias=True),
    )

Ensembling

Number of expetiment	Threshold	Number network from prev. table	Accuracy
1	Without threshold	1, 3, 6 - 12	0.9515
2	> 0.8 = 1	1, 3, 6 - 12	0.9510
3	> 0.5 = 1	1, 3, 6 - 12	0.9484
4	Without threshold	1, 9 - 11	0.9345
5	Without threshold	1, 3, 6 - 12, 14-16	0.9539
6	Without threshold	1, 3, 7 - 11, 14-16	0.9521
7	Without threshold	1, 3, 6 - 12, 14-16, 19-22	0.9577
8	Without threshold	1, 8-11, 15, 16, 19-22	0.9544
9	Without threshold	24, 25	0.9746
10	Without threshold	24-29	0.9771
11	Without threshold	6 TTA of 24	0.9752
12	Without threshold	24-32	0.9783

Types of augmentation

A

# pretrained ImageNet network
T.Resize((224,224))
T.ColorJitter(brightness=0.5, contrast=0.5),
T.RandomRotation((0, 5)),
T.Normalize(mean, std) # ImageNet

B

# pretrained ImageNet network
T.Resize((224,224))
T.ColorJitter(brightness=0.5, contrast=0.5),
T.RandomRotation((-90, 90)),
T.RandomHorizontalFlip(p=0.5),
T.RandomVerticalFlip(p=0.5)
T.Normalize(mean, std) # ImageNet

C

# pretrained ImageNet network
albumentations.Resize(224, 224)
albumentations.RandomRotate90(p=0.5),
albumentations.Transpose(p=0.5),
albumentations.Flip(p=0.5),
albumentations.OneOf([
albumentations.CLAHE(clip_limit=2), albumentations.IAASharpen(), albumentations.IAAEmboss(),
albumentations.RandomBrightness(), albumentations.RandomContrast(),
albumentations.JpegCompression(), albumentations.Blur(), albumentations.GaussNoise()], p=0.5),
albumentations.HueSaturationValue(p=0.5),
albumentations.ShiftScaleRotate(shift_limit=0.15, scale_limit=0.15, rotate_limit=45, p=0.5),
albumentations.Normalize(),