Crossover based technique for data augmentation.

Area of Science:

• Medical Image Analysis
• Artificial Intelligence
• Machine Learning

Background:

• Medical image classification is often limited by small datasets.
• Traditional data augmentation uses label-preserving transformations.
• Non-linear, non-label preserving methods show promise for data enrichment.

Purpose of the Study:

• Propose a novel non-linear data augmentation technique for medical imaging.
• Evaluate the effectiveness of the proposed technique on various datasets and architectures.
• Address the challenge of limited data availability in medical AI applications.

Main Methods:

• Introduce the "Crossover technique" for Convolutional Neural Networks (CNNs).
• Synthesize new training samples by applying two-point crossover on existing images.
• Generate N new samples from N training samples, creating a larger dataset.

Main Results:

• Tested on three public medical datasets (mammograms, skin cancer, brain tumors).
• Improved accuracy across varied network architectures (e.g., VGG-16, VGG-19).
• Achieved notable accuracy gains: 1.47% on mammograms, 3.57% on skin cancer, 0.40% on brain tumors.

Conclusions:

• The crossover technique is a simple yet effective non-linear data augmentation method.
• Successfully tackles limitations of small datasets and class imbalance in medical image analysis.
• The proposed method offers a valuable tool for improving CNN performance in the medical domain.