Purpose: Lesion segmentation is critical for clinicians to accurately stage the disease and determine treatment strategy. Automatic segmentation using deep learning method can improve both the lesion segmentation efficiency and accuracy. However, training a robust deep learning segmentation model requires enough training samples, which may be impractical in many clinical scenarios. This study is to develop a deep learning framework to generate synthetic lesions that can be used together with real lesions for network training.
Methods: The lesion synthesis network is a modified generative adversarial network (GAN). Specifically, we innovated a partial convolution strategy to construct a Unet-like generator. The discriminator is designed using Wasserstein GAN with gradient penalty and spectral normalization. A mask generation method based on Principal component analysis (PCA) was developed to model various lesion shapes. The generated masks are then converted to liver lesions through the lesion synthesis network. The lesion synthesis framework was evaluated with lesion textures and the synthetic lesions were used to train a lesion segmentation network to further validate the effectiveness of the lesion synthetic framework. All the networks are trained and tested on the public dataset from LITS.
Results: The synthetic lesions generated by our approach outperformed the other synthesis approaches by a large margin on image quality and texture evaluation metrics. SSIM and PSNR of the synthetic lesions reached 0.99 and 39.83, and the Kullback-Leibler divergence of GLCM-energy and GLCM-correlation were 0.01 and 0.10. Including the synthetic lesions in the segmentation network improved the segmentation dice performance from 0.671 to 0.712 without augmentation, and 0.706 to 0.729 with augmentation.
Conclusion: The proposed lesion synthesis approach can be used for free-form lesion generation to produce additional labeled training samples. The synthetic data significantly improves the segmentation performance. The approach shows great potential for alleviating the "data paucity" problem.
Funding Support, Disclosures, and Conflict of Interest: Fundamental Research Funds for the Central Universities (Grant No. WK2030000037); Anhui Provincial-level S&T Megaprojects (Grant No. BJ2030480006).