Towards Deep-Learning Partial Volume Correction for SPECT

Théo Kaprélian (1) , Ane Etxebeste (1) , David Sarrut (1)

1 Imagerie Tomographique et Radiothérapie

Partial Volume Effect impacts the spatial resolution of SPECT images. We investigated the feasibility of a deep learning based Partial Volume Correction method (PVCNet) that compensates for the effect of collimator blurring on 2D projections, before reconstruction. A large dataset containing 600,000 pairs of synthetic projections was generated and used to train two consecutive UNets (one for denoising, one for PVC). Scatter and attenuation were not yet considered in the database. Our proposed PVCNet method achieves 12.8% NMAE reduction compared to conventional Resolution Modeling on the IEC phantom but Recovery Coefficients were not always better for smallest spheres.

Mots clés: SPECT Imaging; Deep Learning; Partial Volume Correction; Image Reconstruction

Domaines: Imagerie médicale; Intelligence artificielle [cs.AI]; Traitement du signal et de l'image [eess.SP]; Médecine nucléaire

Théo Kaprélian, Ane Etxebeste, David Sarrut. Towards Deep-Learning Partial Volume Correction for SPECT. 17th International Meeting on Fully 3D Image Reconstruction in Radiology and Nuclear Medicine, 2023. ⟨hal-04590322⟩