Synthesis of Three-dimensional (3D) Multi-contrast Brain Tumor Images with Controlled Tumor Properties Using Generative Models

Motivation: Existing tumor MR image synthetic techniques are implemented for two-dimensional (2D) images and lack continuity in three-dimensional (3D) MR images. Goal(s): To synthesize multi-contrast 3D tumor MRI brain data with controlled tumor properties. Approach: We propose a new 3D model to synthesize tumor images which ensures the continuity of slices. In the model, tumor properties could be controlled by specifying 3D ellipsoids where tumors were expected to be generated. Results: Experimental results demonstrated that the proposed method achieved higher Fréchet inception distance (FID) than an existing 2D method, and synthesized more realistic tumor images. Impact: From non-tumor MR images, images with tumors at controlled locations can be generated using 3D ellipsoids as desirable regions of the tumors. Potential applications include training of machine learning models.