R2U-Net : Medical Image Segmentation

Introduction

In this repository, we implement the R2U-Net network with the Paddle framework. Our model achieves F1-score 0.8232 on DRIVE dataset.

• Original Paper : Recurrent Residual Convolutional NeuralNetwork based on U-Net (R2U-Net) for Medical Image Segmentation
• Dataset : DRIVE

R2U-Net

The proposed models utilize the power of U-Net, Residual Network, as well as RCNN. There are several advantages of these proposed architectures for segmentation tasks. First, a residual unit helps when training deep architecture. Second, feature accumulation with recurrent residual convolutional layers ensures better feature representation for segmentation tasks. Third, it allows us to design better U-Net architecture with same number of network parameters with better performance for medical image segmentation.

Experiment Results

	F1-score	sensitivity	specificity	accuracy	AUC
Original Paper's Results	0.8171	0.7792	0.9813	0.9556	0.9782
Ours Results	0.8232	0.8164	0.9763	0.9557	0.8963

Reprod log

• forward_diff.log
• metric_diff.log
• loss_diff.log
• bp_align_diff.log
• train_align_diff.log
• train log

Train & Test

To train the model yourself, run :

python main.py --model R2U-Net --mode train 

To test the results with the model we provided :

python main.py --model R2U-Net --mode test

Other Parameters:
--dataset_path : path to dataset
--result_path : path to save results
--epoch : training epochs
--batch_size: batch size
--lr : learning rate
--show : show the testing results (default: False)

AI studio link

• https://aistudio.baidu.com/aistudio/projectdetail/2563854