WSSS4LUAD

baseline

https://arxiv.org/pdf/2110.08048.pdf

data

for the training set, the average size of images is (224, 224).

project structure

├─network # backbone models
│  └─structures # scalenet structures
├─result # images and logs for experiment results
├─train
└─utils	

model design

1.backbone [network/wide_resnet_cam.py, network/scalenet_cam.py]

1. add wide resnet-38 for fair comparision (verify the next modifications based on it, and apply scalenet after all settings are ready)

2. multi-scale feature map fusion (concat layer2, layer3, layer4 before fc)

2.training settings: [train.py]

require modification!

classification

• resolution of the patches is 224 × 224
• batch size is set to 20
• training epochs is set to 20
• data augmentation: random horizontal and vertical flip with the probability 0.5.
• learning rate of 1e − 2 with a polynomial decay policy

segmentation phase

• training epochs 20
• learning rate 7e − 2
• no restriction of the image resolution
• data augmentation: horizontal and vertical flip, Gaussian blur and normalization

3.dataloader: [dataset.py (TrainSet, TestSet) ]

1. train dataset: use original pathes without further sub crop
2. train crop: add random resized crop (scale set to (0.25,1), defualt is (0.08, 1))
3. train augmentation: random augmentation for pathology
4. multi-scale test dataset: online crop for test, resize for multiple times firstly (eg, 0.5, 0.75, 1, 1.5, 2), then crop each reiszed image with stride (crop size same to train crop size)

4.test and eval: [test.py, eval.py]

1. test with multi-scale test
2. use x^y for the normal channel (1 - pos.max()) to control the foreground activation scale, and then apply argmax to get psuedo-mask, for the value of y, use grid search based on validation gt

5.improvements: (train.py)

1. norm: mean=[0.678,0.505,0.735] std=[0.144,0.208,0.174]
2. large model: resnest269
3. label balance: use a possitive weight in BCE loss (eg: pos_w = (neg_number / pos_number) ^ 0.5)
4. data synthesis: 4.1 original cutmix in batch; 4.2 cutmix based on label distribution (mainly stroma and tumor); 4.3 cutmix to balance different labels; 4.4 mosaic mix (eg.32x32x64 in seg, 56x56x16 cls, make sure 7 mixed types are balanced)
5. generate pseudo mask without model for single label patches (require corrosion and smoothing )
6. activation drop out: randomly drop high activation
7. area regression loss for single label patches / mixed patches in clssification (top2 loss_area)
8. diff loss weights for single-label / multi-label / mixed label
9. contrastive loss (reduce cosine distance for same category and enlarge it for the different, top2 loss_conl)
10. post-process: 10.1 drop catogory whose area < 5% in subpatch pseudo-mask generation (top3); 10.2 after bg mask, use knn to indentify small area pixels (top1); 10.3 rm small tumer and stroma in normal (top1)