A partial convolution generative adversarial network for lesion synthesis and enhanced liver tumor segmentation.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Computer Vision

Background:

• Accurate lesion segmentation is crucial for disease staging and treatment planning.
• Deep learning models offer improved efficiency and accuracy in medical image segmentation.
• Robust deep learning requires diverse training data, which is often limited for lesions.

Purpose of the Study:

• To develop a deep learning framework for generating synthetic liver lesions with varied locations and sizes.
• To enhance the performance of deep learning-based lesion segmentation models by augmenting training data with synthetic lesions.

Main Methods:

• A modified generative adversarial network (GAN) with a U-Net-like generator utilizing partial convolutions was developed.
• A discriminator incorporating Wasserstein GAN with gradient penalty and spectral normalization was employed.
• Principal Component Analysis (PCA) was used for mask generation to model diverse lesion shapes, which were then synthesized into liver lesions.

Main Results:

• Generated synthetic lesions exhibited similar texture distributions (GLCM-energy, GLCM-correlation) to real lesions.
• Inclusion of synthetic lesions improved segmentation Dice performance from 67.3% to 71.4%.
• Precision and sensitivity for lesion segmentation increased from 74.6% to 76.0% and 66.1% to 70.9%, respectively.

Conclusions:

• The proposed lesion synthesis framework effectively generates realistic synthetic liver lesions.
• Augmenting training datasets with these synthetic lesions significantly enhances the performance of deep learning segmentation models.
• This approach offers a viable solution to data scarcity challenges in training robust medical image segmentation models.