Click here to


Are you sure ?

Yes, do it No, cancel

Realistic Lesion Generation Using Generative Adversarial Networks and Radiomics Supervision

S Pan*, J Stayman, C Lin, G Gang, Johns Hopkins University, Baltimore, MD


(Sunday, 7/12/2020) 3:30 PM - 4:30 PM [Eastern Time (GMT-4)]

Room: Track 1

Purpose: Realistic lesion simulation is widely used to provide targets in virtual clinical trials and as imaging tasks in system assessment and optimization. Traditional lesion generation primarily replied on procedural approaches. In this work, we leverage large public databases with patient lesions from lung CT scans and applied a generative neural network for realistic lesion simulation.

Methods: We implemented a Wasserstein generative adversarial network (GAN) with gradient penalty. Novel to our implementation, we included an intermediate supervision step and directly used radiomics features to supervise an intermediate generator layers. For initial implementation, we used Gray-Level Co-Occurrence Matrix (GLCM) homogeneity. We trained the network on two categories of lesions, solid and non-solid (including ground glass and part-solid), based on contours and annotations provided in the Lung Image Database Consortium (LIDC) database. We evaluated the realism of the generated lesions by comparing quantitative measures of texture and shape within the training group and generated group.

Results: The network is capable of generating lesions of each category. Within the non-solid category, the network can produce a variety of textures (part-solid, calcification). The network also demonstrates ability to generate lesions of different shapes as well as spiculated and smooth boundaries. The distribution of quantitative texture and shape features agree well for most features as indicated by a low KL-divergence score.

Conclusion: This work presents a novel implementation of GAN capable of generating realistic new lesions based on existing lesions database and learning the statistical distribution of lesion features based on radiologists’ annotations. This method can be used to for big data generation required for image quality assessment.

Funding Support, Disclosures, and Conflict of Interest: NIH R21CA219608


Not Applicable / None Entered.


IM- CT: Phantoms - digital

Contact Email