Improving CNN Training using Disentanglement for Liver Lesion Classification in CT.

Area of Science:

• Medical Image Analysis
• Artificial Intelligence
• Computer Vision

Background:

• High-quality training data is crucial for medical image analysis algorithms.
• Data scarcity often limits the performance of neural networks in this field.
• Recent advances in image generation offer potential solutions using synthetic data.

Purpose of the Study:

• To develop a novel method for synthesizing diverse medical image data.
• To improve the accuracy of medical image analysis algorithms by addressing data limitations.
• To explore the use of disentangled representations for controlled synthetic data generation.

Main Methods:

• Separating key appearance factors in training data using a disentanglement-based scheme.
• Synthesizing new training samples by mixing specified and unspecified representations.
• Applying the method to liver lesion classification tasks using CT images.

Main Results:

• The proposed data synthesis method demonstrated an average accuracy improvement of 7.4% in liver lesion classification compared to baseline training.
• The approach effectively generated varied synthetic samples for specific classes.
• Controlled generation of appearance features was achieved through disentangled representations.

Conclusions:

• Data augmentation through controlled synthetic sample generation can significantly enhance medical image analysis.
• Disentanglement-based schemes offer a promising approach for creating effective synthetic training data.
• This method provides a viable strategy to overcome data bottlenecks in developing robust medical imaging algorithms.