Adaptive Evolutionary Optimization of Deep Learning Architectures for Focused Liver Ultrasound Image Segmentation.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Biomedical Engineering

Background:

• Liver ultrasound segmentation is difficult due to poor image quality and variability.
• Standard deep learning (DL) models may not be optimal for specific liver ultrasound segmentation tasks.
• Accurate segmentation is crucial for quantitative ultrasound (QUS) liver fat measurement.

Purpose of the Study:

• To develop a generalizable framework for optimizing DL models for liver ultrasound segmentation.
• To enhance the accuracy of segmentation for quantitative ultrasound applications.
• To adapt DL architectures for improved performance on challenging liver ultrasound images.

Main Methods:

• A generalizable framework utilizing an adaptive evolutionary genetic algorithm to optimize U-Net DL models.
• Simultaneous adjustment of network depth, width, dropout, and skip connections.
• Evaluation of various U-Net configurations based on segmentation performance for liver ultrasound.

Main Results:

• The optimized U-Net model with depth 4 and filter sizes [16, 64, 128, 256] achieved the highest mean adjusted Dice score of 0.921.
• The adaptive evolutionary optimization significantly outperformed other configurations.
• Three-fold cross-validation with early stoppage was employed to validate the results.

Conclusions:

• Adaptive evolutionary optimization effectively enhances DL architectures for liver ultrasound segmentation.
• The optimized model shows significant potential for improving liver fat measurement using QUS.
• Future research can extend this optimization approach to other imaging modalities and DL architectures.