Recurrent Residual Convolutional Neural Network based on U-Net (R2U-Net) for Medical Image Segmentation (1802.06955v5)

Abstract: Deep learning (DL) based semantic segmentation methods have been providing state-of-the-art performance in the last few years. More specifically, these techniques have been successfully applied to medical image classification, segmentation, and detection tasks. One deep learning technique, U-Net, has become one of the most popular for these applications. In this paper, we propose a Recurrent Convolutional Neural Network (RCNN) based on U-Net as well as a Recurrent Residual Convolutional Neural Network (RRCNN) based on U-Net models, which are named RU-Net and R2U-Net respectively. 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. The proposed models are tested on three benchmark datasets such as blood vessel segmentation in retina images, skin cancer segmentation, and lung lesion segmentation. The experimental results show superior performance on segmentation tasks compared to equivalent models including U-Net and residual U-Net (ResU-Net).

• The paper presents RU-Net and R2U-Net, which incorporate recurrent layers and residual connections to improve segmentation performance.
• It demonstrates superior results across retina, skin, and lung image datasets with high accuracy and AUC metrics.
• The models achieve enhanced feature propagation and boundary delineation without extra computational cost, promising real-time diagnostic applications.

Recurrent Residual Convolutional Neural Network based on U-Net (R2U-Net) for Medical Image Segmentation

The paper under review proposes enhancements to the U-Net architecture, leveraging Recurrent Convolutional Neural Networks (RCNN) and Recurrent Residual Convolutional Neural Networks (RRCNN), introduced as RU-Net and R2U-Net, respectively. These novel architectures demonstrate improved performance for medical image segmentation tasks, validated through comprehensive experiments on three crucial datasets: blood vessel segmentation in retina images, skin cancer segmentation, and lung lesion segmentation.

Architectural Enhancements

Fundamentals of U-Net

U-Net's success is predicated on its encoder-decoder structure, which captures essential features through convolutional layers and refines segmentation maps via deconvolutional layers. However, the authors aim to surmount inherent limitations in U-Net by integrating residual and recurrent layers.

Proposed Models: RU-Net and R2U-Net

• RU-Net: This model implements recurrent convolution layers (RCLs) within the U-Net framework to better capture dependencies and hierarchical features through multiple time steps.
• R2U-Net: Extends RU-Net by incorporating residual connections, which aid in mitigating the vanishing gradient problem and further enhance feature propagation across layers.

Both models preserve the original parameter count of U-Net, indicating computational efficiency with more robust performance.

Experimental Validation

Datasets and Implementation

1. Retina Blood Vessel Segmentation: Evaluated on DRIVE, STARE, and CHASE_DB1 datasets. The DRIVE dataset involved 20 samples for training and testing each, with a processing approach that included patch extraction (48x48 pixels).
2. Skin Cancer Segmentation: Data from the 2017 ISIC challenge was used, consisting of 2000 images, with preprocessing steps reducing image sizes to 256x256 pixels.
3. Lung Segmentation: Leveraged a subset from the LUNA challenge with 534 2D images, resized to 256x256 pixels.

Performance Metrics

The paper utilizes accuracy (AC), sensitivity (SE), specificity (SP), Dice coefficient (DC), and Jaccard similarity (JS) to assess segmentation quality. Key findings include:

• DRIVE Dataset: R2U-Net achieved an accuracy of 0.9556 and AUC of 0.9784, outperforming standard and residual U-Net models.
• STARE Dataset: Statistically significant improvements were observed with R2U-Net, marked by an accuracy of 0.9712 and AUC of 0.9914.
• CHASE_DB1 Dataset: The model excelled with an accuracy of 0.9634 and AUC of 0.9815.
• Skin Cancer and Lung Segmentation: R2U-Net consistently provided superior results with higher DC and AC, affirming its capability in diverse medical imaging tasks.

Implications and Future Directions

Practical and Theoretical Contributions

These models' ability to deliver enhanced segmentation results without increasing computational overhead has profound implications for real-time medical diagnostics. The recurrent and residual integrations facilitate better feature accumulation and boundary delineation, which are crucial for accurate medical image analysis.

Future Developments

The demonstrated efficacy of RU-Net and R2U-Net invites further exploration into:

1. Multi-Modal Medical Imaging: Extending the application to other imaging modalities, such as MRI and CT scans.
2. Feature Fusion Strategies: Enhancing feature fusion techniques between encoding and decoding layers to further improve model performance.
3. 3D Segmentation Tasks: Evaluating the models in 3D medical image segmentation to verify their robustness and scalability.

In conclusion, the RU-Net and R2U-Net models represent significant advancements in the segmentation of medical images, combining the strengths of recurrent and residual learning within the U-Net architecture. This innovative approach secures prominent gains in prediction accuracy and robustness across various medical imaging datasets, rendering it an invaluable contribution to the field of computer-aided medical diagnosis.