Incorporating prior shape knowledge via data-driven loss model to improve 3D liver segmentation in deep CNNs

Area of Science:

• Medical Imaging
• Computer Vision
• Artificial Intelligence

Background:

• Convolutional neural networks (CNNs) show great promise for liver segmentation but face challenges with low-contrast images and variations in liver shape and appearance.
• Integrating prior knowledge into deep CNNs can enhance model performance and generalization for medical image analysis.

Purpose of the Study:

• To improve 3D liver segmentation using deep CNNs by incorporating prior shape knowledge.
• To enhance the generalization and robustness of liver segmentation models.

Main Methods:

• A convolutional denoising auto-encoder was used to learn global 3D liver shape information in a latent space.
• This learned knowledge was integrated into a hybrid loss function combined with Dice loss for a 3D U-Net segmentation model.
• The hybrid model was trained to leverage global shape priors while optimizing segmentation accuracy.

Main Results:

• The proposed training strategy achieved a Dice score of 97.62% on the Sliver07-I liver dataset, competitive with state-of-the-art methods.
• The hybrid model demonstrated improved generalization and robustness, outperforming the baseline 3D U-Net on unseen images.
• The incorporation of data-driven prior shape knowledge enhanced segmentation performance.

Conclusions:

• Prior shape knowledge significantly boosts the performance of deep CNNs in liver segmentation tasks.
• The proposed method enhances model generalization and robustness through abstract features learned via a data-driven loss model.