Robust Lesion Segmentation on Emission Computed Tomography with nmODE

Nuclear medicine imaging plays a pivotal role in the diagnosis of systemic bone diseases. However, conventional discrete models struggle to handle the inherent fuzziness and randomness of such images. Recently proposed, the neural memory ordinary differential equation (nmODE) model exhibits intriguing characteristics and has shown significant improvements in classification tasks. In this study, we leverage the powerful nonlinear and dynamic modeling capabilities of nmODE and extend its application to the field of nuclear medicine image segmentation. The experimental results affirm that the inherent stability of nmODE is highly suitable for addressing the task of lesion segmentation in nuclear medicine images characterized by indistinctness and stochasticity.