DyABD: A Dataset and Technique for Synthetically Generating Dynamic Abdominal MRIs with Dual Class and Anatomically Conditioned Diffusion Models

Niamh Belton, Victoria Joppin, Aonghus Lawlor, Kathleen M Curran, Catherine Masson, Thierry Bege, David Bendahan

Keywords: Diffusion, Dynamic, MRI, Abdominal, Generative, Deep Learning, Medical Imaging, Image Synthesis

Abstract: An abdominal hernia is a protrusion of intestine or tissue in the abdominal wall and is known to cause debilitating pain. The recurrence rate of abdominal hernia varies from 30\% to 80\%, meaning it is paramount to improve our understanding of the mechanical functionality and physiology of the abdominal wall. This work proposes DyABD, a dataset of dynamic abdominal MRIs (2D+t) of hernia patients and a 3D dual class and anatomically conditioned Denoising Diffusion Probabilistic Model (DDPM) that can perform the unique task of synthesising hernia patients performing any of three exercises, breathing, coughing or a Valsalva maneuver, whilst also taking into account whether the patient is pre or post corrective abdominal surgery. DyABD requires a subject prior as input which consists of the first 2D slice of the dynamic MRI sequence and the associated abdominal muscle masks of the first 2D slice which ensures anatomical correctness is preserved during synthesis. This work is based on 121 dynamic MRI volumes which will be made available for sharing as part of the complete dataset of approximately $300$ volumes. The preliminary results of DyABD demonstrate its ability to model the mechanical functionality of the abdominal wall. Examples of generated volumes are made available at https://github.com/niamhbelton/DyABD/.

Submission Number: 166