Evolutionary Deep Attention Convolutional Neural Networks for 2D and 3D Medical Image Segmentation

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Computational Neuroscience

Background:

• Medical image segmentation using deep convolutional neural networks (CNNs) faces challenges due to limited labeled data and computational constraints.
• Incorporating attention mechanisms in CNNs is crucial for improving segmentation accuracy despite image noise and diversity, but designing such complex networks is difficult.
• Neuroevolution offers automated network design but is computationally intensive, particularly for 3D network architectures.

Purpose of the Study:

• To develop an automatic, efficient, accurate, and robust technique for creating deep attention CNNs for 2D and 3D medical image segmentation.
• To address the computational expense of neuroevolution for complex network structures and 3D applications.
• To optimize the integration of attention modules within a U-Net-based architecture for enhanced spatial information recovery.

Main Methods:

• An automated technique utilizing neuroevolution to design deep attention CNNs for medical image segmentation.
• The evolutionary approach identifies optimal combinations of six attention modules to improve spatial information transfer in a U-Net architecture.
• Evaluation performed on six diverse CT and MRI datasets for both 2D and 3D image segmentation tasks.

Main Results:

• The proposed neuroevolutionary technique successfully developed deep attention CNNs that significantly improved medical image segmentation.
• The model demonstrated superior performance in recovering and transferring spatial information between the downsampling and upsampling paths of the U-Net network.
• Comparative analysis showed the proposed model outperformed all state-of-the-art manual and automatically designed segmentation models across all tested datasets.

Conclusions:

• The presented neuroevolutionary approach provides an efficient and effective method for automatically designing high-performance deep attention CNNs for medical image segmentation.
• This technique overcomes the limitations of manual design and computationally expensive neuroevolution for complex 3D networks.
• The model's ability to outperform existing methods highlights its potential for advancing medical image analysis and segmentation accuracy.