Image Augmentation based on Variational Autoencoder for Breast Tumor Segmentation.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Oncology

Background:

• Manual segmentation of breast tumors from Dynamic Contrast-Enhanced Magnetic Resonance Imaging (DCE-MRI) is labor-intensive, prone to errors, and suffers from inter-observer variability.
• Accurate tumor segmentation is crucial for computable radiomics analysis in breast cancer.
• Deep learning models have shown promise in automating image segmentation tasks.

Purpose of the Study:

• To develop and evaluate a novel deep learning model for automated 3D breast tumor segmentation from DCE-MRI.
• To address challenges of limited training data by incorporating a variational auto-encoder.

Main Methods:

• A 3D Connected-UNets architecture based on an encoder-decoder framework was employed.
• A variational auto-encoder was integrated to augment the training data and enhance feature learning.
• Post-processing steps included a fully connected 3D conditional random field and 3D connected module evaluation to refine segmentation boundaries and reduce noise.

Main Results:

• The proposed 3D Connected-UNets model demonstrated superior performance in breast tumor segmentation.
• Evaluated on public (INbreast, DDSM) and private datasets, the model consistently outperformed existing state-of-the-art methods.
• The integration of variational auto-encoder and post-processing techniques improved segmentation accuracy.

Conclusions:

• The 3D Connected-UNets model offers a robust and accurate solution for automated breast tumor segmentation in DCE-MRI.
• This approach has the potential to significantly aid in radiomics analysis and improve breast cancer diagnostics.
• The method effectively handles limited dataset sizes and enhances segmentation quality.