Lung Giants: An Evaluation Framework for Mask-Conditioned Lung CT Synthesis

Abstract

As it stands, the development of robust AI-based diagnosis aid systems is greatly hindered by the lack of access to high-quality and fully annotated medical imaging data, which is both scarce, hard to acquire and subject to numerous privacy regulations. This issue has mostly been solved through advancements in the GenAI field, although within the study of lung cancer specifically, two issues still remain: non-conditional synthesis, despite successful, is unable to fill more specific gaps in available datasets; and traditional variable-conditioned generation is much too demanding and irregular to be considered viable for every occasion. As such, this work presents an evaluation framework for up-to-date SotA mask-conditioned 3D lung CT synthesis methods, in order to evaluate performance across various metrics and visual aspects of the artificial data. Experiments show that despite successful in creating visually realistic and diverse samples, the explored methods do have limitations, namely when dealing with smaller nodule sizes and some internal texture. Though not entirely uniform, evaluation of such published techniques allow us to form a solid base upon which to build future work on 3D mask-conditioned lung CT generation. In sum, this work poses itself as a preliminary study of current SotA to further fundament an upcoming more comprehensive 3D mask-conditioned lung CT generative technique.