Multi-objective data enhancement for deep learning-based ultrasound analysis.

Area of Science:

• Medical Informatics
• Artificial Intelligence in Medicine

Background:

• Deep learning models for treatment recommendations struggle with small medical datasets, leading to overfitting and poor performance.
• Automatic generation of treatment recommendations using deep learning is an area of significant research interest.

Purpose of the Study:

• To propose a multi-objective data enhancement method to address overfitting in deep learning models for medical data.
• To improve the quantity and quality of treatment recommendations generated from limited datasets.

Main Methods:

• Defined a main task and several auxiliary tasks using the same dataset to learn diverse knowledge.
• Employed a Soft Parameter Sharing method to transfer knowledge between models trained on different tasks.
• Utilized a thyroid nodule ultrasound dataset with professional doctor-labeled findings, impressions, and treatment recommendations.

Main Results:

• The proposed multi-objective data enhancement method demonstrated superior performance compared to existing methods.
• The approach effectively mitigates overfitting by leveraging knowledge from auxiliary tasks.
• Improved generation of high-quantity treatment recommendations was achieved.

Conclusions:

• Multi-objective data enhancement with soft parameter sharing is a viable strategy for improving deep learning performance on small medical datasets.
• This method enhances the generation of accurate and abundant treatment recommendations.
• The approach holds promise for advancing AI-driven clinical decision support systems.