Seamless Lesion Insertion for Data Augmentation in CAD Training.

Area of Science:

• Medical Imaging
• Computer Vision
• Machine Learning

Background:

• Classifier performance relies heavily on training data size and representativeness.
• Acquiring and labeling medical data can be challenging and costly.
• Data augmentation offers a potential solution for limited datasets in medical applications.

Purpose of the Study:

• To evaluate the effectiveness of a lesion insertion tool for creating synthetic chest CT data.
• To investigate the impact of augmenting training sets with synthetic data on lung nodule detection classifiers.
• To determine if this augmentation method can improve performance with small training datasets.

Main Methods:

• Developed an image blending tool for seamless lesion insertion into target images.
• Applied various transformations to lesions before insertion.
• Created realistic synthetic lung nodule samples in chest CT images.
• Augmented training sets of varying sizes with synthetic samples.
• Trained and evaluated classifiers for lung nodule detection.

Main Results:

• The lesion insertion method successfully generated realistic synthetic chest CT samples.
• Augmenting small training datasets with synthetic samples improved classifier performance.
• The proposed method demonstrated a reduction in the need for extensive actual patient data.

Conclusions:

• The developed lesion insertion technique is effective for data augmentation in medical imaging.
• Synthetic data generation can significantly enhance classifier performance, especially for limited datasets.
• This approach offers a viable strategy to mitigate challenges in medical data acquisition and labeling.