I. DATASET AND KAGGLE COMPETITION

Link to challenge: https://www.kaggle.com/c/bkai-igh-neopolyp/

This dataset contains 1200 images (1000 WLI images and 200 FICE images) with fine-grained segmentation annotations. The training set consists of 1000 images, and the test set consists of 200 images. All polyps are classified into neoplastic or non-neoplastic classes denoted by red and green colors, respectively. This dataset is a part of a bigger dataset called NeoPolyp.

II CODE STRUCTURE

The code is structured as following:

project
│   README.md
└───data
│   └───test
|	|	|	test1.jpeg
|	|	|	test2.jpeg
|	|	|	...
│	│
│   └───train 
│	│
│   └───train_gt 
└───logs
│
└───models 
│
└───loss 
|
└───res 
|
└───weights
|	config.py
|	dataset.py
|	main.py
| 	mask2csv.py 
|	test.py

data : Where you put the training data and ground truth. There are three sub-folders test, train and train_gt in this folder. Please download data from https://www.kaggle.com/c/bkai-igh-neopolyp/data , and put the images in the corresponding folders ( only the images).

logs: You should save the training log of the models here.\

models : Code implementation of the models. Current supported models are: Unet and Attention Unet , and with three diferent backbones: Mobilnetv2,Efficientnet B0 and VGG-16.

res : inference results are saved here.

loss : implementation of loss functions

weights : weights are saved here.

config.py : Config for paths and model parameters

dataset.py : dataset-related utility functions

main.py : main training script.

mask2csv : convert from image to csv for kaggle submission

III. Training

To train the model for polyps segmentation , please change the following parameters in config.py

IMAGE_PATH  =  '/media/HDD/bkai/data/train'
MASK_PATH  =  '/media/HDD/bkai/data/train_gt'
WIDTH  =  256
HEIGHT  =  256
BATCH_SIZE  =  8
MODEL_SELECTION  =  "mb2_unet"

IMAGE_PATH : path to the train dataset directory MASK_PATH : path to the train ground truth dataset directory TEST_PATH : path to the test dataset directory WIDTH and HEIGHT : input size of the model. Currently only supports 256x256. You can make it dynamic by making modifications to models in the folder models. BATCH_SIZE: batch size MODEL_SELECTION: Currently supports: unet (Unet with VGG-16 backbone), unet_eff0 (Unet with Efficientnet B0 backbone),mb2_unet (Unet with MobilenetV2 backbone) ,att_unet (Attention Unet), att_unet_mb2 and unet_att_eff0.

Then simply run main.py

IV. INFERENCE

To run inference , please change the following parameters in config.py

TEST_PATH  =  '/media/HDD/bkai/data/test'
WEIGHT_PATH  =  "/media/HDD/bkai/weights/mobinetv2_unet.h5"
SAVE_FOLDER  =  "/media/HDD/bkai/res"
SHOW_IMAGE  =  True

TEST_PATH : path to the test dataset directory WEIGHT_PATH : path to the trained weight SAVE_FOLDER : folder to save the inference result SHOW_IMAGE: If you want to visualize the result one by one and show overlay result , please set to True. If you want to save the segmentation result , set to False. Please note that the visualization result if SHOW_IMAGE= True won't be smooth since it's the original one-hot map of size 256x256x3 and has not been resized to the original input size yet.

V. CONVERT TO CSV

To convert result to .csv , please change the following parameters in config.py

MASK_RES_PATH  =  '/media/HDD/bkai/res'

MASK_RES_PATH : path to the inference result directory

VI. DETAILED EXPLAINATION

Unet

Information about Unet models are located at models/Unet.py , models/Unetefficientb0.py, models/Unetmobilev2.py.

Attention Unet

Information about Attention Unet models are located at models/AttentionUnet.py , models/AttentionUnetefficientb0.py, models/AttentionUnetmobilev2.py.

DATASET

We need segent 3 classes:

0 if the pixel is part of the image background (denoted by black color);
1 if the pixel is part of a non-neoplastic polyp (denoted by green color);
2 if the pixel is part of a neoplastic polyp (denoted by red color).

See function convert2TfDataset in dataset.py.

TEST

The output of the model will be of size H,W,3, with 3 channels corresponding to 3 classes. The output will then be converted to corresponding RGB image

colors  =  np.array([[ 0,  0,  0],
					[ 0,  255,  0],
					[ 255,  0,  0]])

p  =  cv2.resize(p.astype(np.float32),(w,h), cv2.INTER_CUBIC) #H,W,3
p  =  np.argmax(p,axis=-1) # 995,1280
rgb  =  np.zeros((*p.shape,  3))
for  label,  color  in  enumerate(colors):
	rgb[p  ==  label] =  color

GivralNguyen / BKAI-IGH-Neopolyp-Segmentation