RA-UNet: A hybrid deep attention-aware network to extract liver and tumor in CT scans (1811.01328v1)

Abstract: Automatic extraction of liver and tumor from CT volumes is a challenging task due to their heterogeneous and diffusive shapes. Recently, 2D and 3D deep convolutional neural networks have become popular in medical image segmentation tasks because of the utilization of large labeled datasets to learn hierarchical features. However, 3D networks have some drawbacks due to their high cost on computational resources. In this paper, we propose a 3D hybrid residual attention-aware segmentation method, named RA-UNet, to precisely extract the liver volume of interests (VOI) and segment tumors from the liver VOI. The proposed network has a basic architecture as a 3D U-Net which extracts contextual information combining low-level feature maps with high-level ones. Attention modules are stacked so that the attention-aware features change adaptively as the network goes "very deep" and this is made possible by residual learning. This is the first work that an attention residual mechanism is used to process medical volumetric images. We evaluated our framework on the public MICCAI 2017 Liver Tumor Segmentation dataset and the 3DIRCADb dataset. The results show that our architecture outperforms other state-of-the-art methods. We also extend our RA-UNet to brain tumor segmentation on the BraTS2018 and BraTS2017 datasets, and the results indicate that RA-UNet achieves good performance on a brain tumor segmentation task as well.

• The paper introduces RA-UNet, a hybrid deep attention-aware network designed to accurately segment liver and tumor regions in CT scans.
• It combines a 3D U-Net framework with residual attention mechanisms to address gradient vanishing and improve computational efficiency.
• Evaluated on LiTS and 3DIRCADb datasets, RA-UNet achieves high Dice scores, outperforming existing state-of-the-art segmentation methods.

RA-UNet: A Hybrid Deep Attention-Aware Network for Liver and Tumor Segmentation in CT Scans

The paper presents a novel approach to the automated extraction of liver and tumor from CT volume using a hybrid deep attention-aware neural network called RA-UNet. This work addresses challenges associated with the heterogeneous and diffusive nature of liver and tumor shapes within 3D medical imaging, which complicate accurate segmentation.

Methodology

The architecture of RA-UNet is grounded in the 3D U-Net framework, incorporating both an attention mechanism and residual learning. This approach allows the network to handle the gradient vanishing problem and increase network depth without a degradation in performance. The attention mechanism enables the network to focus on significant features, enhancing efficiency and reducing computational overhead.

RA-UNet is structured in a three-phase process:

1. Liver Localization: Utilizing a 2D RA-UNet (RA-UNet-{1}) to establish a coarse liver boundary box aimed at reducing computational effort by narrowing the region of interest.
2. Liver Segmentation: A 3D RA-UNet (RA-UNet-{2}) provides precise segmentation using rich 3D contextual information captured within the liver boundary box.
3. Tumor Extraction: Further utilization of RA-UNet-{2} focuses on segmenting tumors within the delineated liver volumes.

Notably, the RA-UNet framework leverages a residual attention mechanism for the first time in medical volumetric image analysis, which optimizes the segmentation by adaptively focusing on informative features while suppressing irrelevant data. The integration of 3D networking further distinguishes RA-UNet by fully utilizing spatial information, which proves beneficial for delineating complex structures such as liver and tumor in CT images.

Evaluation and Results

The methodology was evaluated using the MICCAI 2017 Liver Tumor Segmentation dataset and the 3DIRCADb dataset, with strong performance metrics. RA-UNet achieved a Dice score of 0.961 for liver segmentation and 0.595 for tumor segmentation on the LiTS test dataset, while on the 3DIRCADb dataset, it attained Dice scores of 0.977 for liver and 0.830 for tumor. Such results indicate RA-UNet's ability to outperform existing state-of-the-art methods, demonstrating its robustness and efficiency in clinical datasets.

Furthermore, RA-UNet's adaptability to other medical imaging tasks was corroborated through an extension to brain tumor segmentation using the BraTS2018 dataset, yielding competitive results and asserting the model's potential applicability in various segmentation challenges.

Implications and Future Directions

The introduction of RA-UNet presents substantial implications for both clinical practice and the broader field of medical image analysis. By facilitating precise liver and tumor detection with reduced computational costs, RA-UNet can enhance pre-surgical planning and improve personalized treatment strategies, ultimately contributing to better patient outcomes.

On a broader theoretical level, the successful application of attention mechanisms and residual learning in 3D image segmentation highlights the potential for these strategies to tackle challenges across diverse datasets and tasks. Future research may focus on generalizing RA-UNet's methodology to incorporate other organ systems or imaging modalities, optimizing for computational efficiency, and refining the network's capacity to handle a wider array of volumetric imaging tasks. As AI continues to advance, integrating such sophisticated models into real-world medical platforms could revolutionize diagnostic and treatment paradigms.