AdaEn-Net: An ensemble of adaptive 2D-3D Fully Convolutional Networks for medical image segmentation

Area of Science:

• Medical image analysis
• Deep learning for medical imaging
• Computer-aided diagnosis

Background:

• Fully Convolutional Networks (FCNs) are effective for image segmentation but manual design is time-consuming and leads to complex models.
• Automating the design of efficient 3D medical image segmentation architectures is crucial.
• Existing methods often fail to fully leverage volumetric data or optimize model size.

Purpose of the Study:

• To propose AdaEn-Net, a self-adaptive 2D-3D ensemble of FCNs for 3D medical image segmentation.
• To incorporate volumetric data and optimize both performance and size for specific datasets.
• To automatically search for high-performing and efficient segmentation architectures.

Main Methods:

• Developed AdaEn-Net, an ensemble of a 2D FCN for intra-slice information and a 3D FCN for inter-slice information.
• Employed a multiobjective evolutionary algorithm to optimize architecture and hyperparameters, maximizing segmentation accuracy and minimizing network parameters.
• Evaluated the model on prostate segmentation (PROMISE12) and cardiac segmentation (MICCAI ACDC) challenges.

Main Results:

• AdaEn-Net ranked 9th out of 297 in the PROMISE12 challenge, outperforming an auto-generated network with 13x fewer parameters.
• Achieved a top 8 ranking in the MICCAI ACDC challenge, surpassing a reinforcement learning-designed architecture with 1.25x fewer parameters.
• Demonstrated effective exploitation of volumetric information for efficient and accurate 3D medical image segmentation.

Conclusions:

• AdaEn-Net offers an automated approach to designing efficient and accurate 3D medical image segmentation models.
• The self-adaptive ensemble effectively integrates 2D and 3D convolutional information.
• The evolutionary optimization successfully balances segmentation performance and model complexity.