A generative flow-based model for volumetric data augmentation in 3D deep learning for computed tomographic...

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Deep Learning

Background:

• Deep learning improves computer-aided detection (CADe) in medical imaging.
• CT colonography CADe performance is hindered by limited training data size and variety.
• Effective data augmentation is crucial for training deep learning models in this field.

Purpose of the Study:

• To develop and evaluate a flow-based generative model for 3D data augmentation of colorectal polyps.
• To enhance the training of deep learning models for CADe in CT colonography.
• To address the limitations of small and diverse training datasets.

Main Methods:

• Developed a 3D-convolutional neural network (3D CNN) using a 3D flow-based generative model (3D Glow).
• Trained 3D Glow on a clinical CT colonography dataset to generate synthetic polyp volumes of interest (VOIs).
• Evaluated synthetic data using human observer studies and a CADe-based polyp classification study with a 3D DenseNet.

Main Results:

• Human observers could not statistically distinguish between real and 3D Glow-generated synthetic polyps.
• Polyp classification performance of a 3D DenseNet significantly improved when trained with 3D Glow data augmentation.
• Achieved statistically significant improvements in polyp classification compared to alternative augmentation methods.

Conclusions:

• 3D Glow generates synthetic polyps that are visually indistinguishable from real ones.
• Data augmentation with 3D Glow substantially enhances the performance of 3D CNNs for CADe in CT colonography.
• 3D Glow presents a promising approach for advancing deep learning in CT colonography CADe.