Since the image resolution is quite large,I decide to use iafoss's trick to divide image into titles.By doing this,i can create small dataset without losing information from original image.I use this notebook to do this step

|__converting-to-256x256.ipynb

dataset :https://www.kaggle.com/datasets/bhuynguyn/hubmaphacking2022-comp-256x256

After creating new dataset, I create data training pipeline:
I applied some data augmentation techniques such as CLAHE, RandomBrightness, RandomContrast,… to
enhance the quality of images and enrich training data.
data_prepare
|__dataloader.py

link :https://arxiv.org/pdf/1904.04445.pdf

Authors inserted SCSE modules after each encoder and decoder blocks.They act as attention modules,using spatial attention mechanism helps model to focus on important features of FTU and suppress less important ones.

Additionally, in the bottleneck block between the encoder and the decoder authors use Feature Pyramid Attention module, which increases the receptive field by fusing features from different pyramid scales.By doing this, model will be able to focus on object at different places in the picture.

To exploiting feature maps from different scales authors use Hypercolumns. They stack the upsampled feature maps from all decoder blocks and use them as the input to the final layer. Therefore,they could get more precise localization and captures the semantics at the same time. (I will public my training later)

In this architecture, backbone resnet50 is used for extracting feature maps from different convolution layers.Feature maps after each convolution layer are fed into RFB(receptive field block) .RFB block is designed with 4 branches,using diliate convolution layers with different rate.With this design,this module not only capture global contextual information of tissues in the picture, but it also reduces the computation costs.The output map of RFB module is fed into RA(reverse attention) module to capture contextual informations and use them as a guild to combine with feature map of each layer to have both spatial informations and contextual ones.

optimizer :Adam
min_lr=1e-07
max_lr=1e-04
epoch :168 epochs
CosineAnnealingLR
Result: https://www.kaggle.com/code/bhuynguyn/training-pranet/notebook