Efficient In-Training Adaptive Compound Loss Function Contribution Control for Medical Image Segmentation.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Computer Vision

Background:

• Image segmentation is vital for clinical applications like disease diagnosis and monitoring.
• Deep neural networks are state-of-the-art for segmentation but face challenges like class imbalance.
• Compound loss functions, combining binary cross-entropy (BCE) and Dice loss, are common but require extensive tuning.

Purpose of the Study:

• To develop an efficient method for training deep neural networks for image segmentation.
• To address the challenge of class imbalance in medical image segmentation.
• To eliminate the need for tedious hyperparameter fine-tuning in compound loss functions.

Main Methods:

• Proposed an adaptive approach to dynamically control the contribution of individual loss functions during training.
• Integrated binary cross-entropy (BCE) and Dice loss adaptively.
• Focused on improving segmentation model precision and recall without manual hyperparameter optimization.

Main Results:

• The adaptive method eliminates the need for multiple fine-tuning iterations.
• Achieved desired precision and recall for segmentation models more efficiently.
• Reduced time and energy consumption associated with model training.

Conclusions:

• The proposed adaptive loss function control is an efficient solution for class imbalance in medical image segmentation.
• This approach streamlines the training process for deep learning models.
• Enables more accessible and resource-efficient development of clinical image analysis tools.