NUCLIDE: A Novel Unsupervised Clustering-based Image Denoising and Enhancement for 4D Dynamic PET/MRI Data

Motivation: Dynamic PET denoising through deep learning struggles with contrast preservation and activity fidelity, compounded by black-box models and the absence of paired noisy-clean PET data. Current methods often oversmooth, obscuring crucial details. Goal(s): We propose an unsupervised 4D PET denoising approach that enhances image quality while maintaining TACs, leveraging the clustering of spatiotemporal features. Approach: Our method utilizes K-means clustering on temporal PCA, spatial MRI, and anatomical features to create localized STC maps. Lowpass filtering is applied in a Radon-transformed feature space for targeted denoising without blurring. Results: Results show improved SNR & CNR, with an error reduction of 54.2% on simulated OSEM data. Impact: Our method enables robust, unsupervised denoising for PET/MRI, preserving critical TACs and structural information. This method has applications in clinical settings and is adaptable to multimodal imaging.