Shorter SPECT scans using self-supervised coordinate learning to synthesize skipped projection views

Area of Science:

• Nuclear Medicine
• Medical Imaging
• Artificial Intelligence in Healthcare

Background:

• Extended SPECT imaging duration, especially under low-count conditions like Lu-177 SPECT, poses clinical challenges.
• Reducing scan times is crucial for patient comfort and improving throughput in nuclear medicine.
• Current methods for reducing SPECT acquisition time may compromise image quality and quantitative accuracy.

Purpose of the Study:

• To develop and evaluate a self-supervised learning approach (SpeRF) for synthesizing SPECT projection views.
• To shorten SPECT scan times in clinical settings, particularly for Lu-177 imaging, by reducing the number of acquired projections.
• To maintain quantitative accuracy and image quality despite reduced acquisition duration.

Main Methods:

• Developed SpeRF, a SPECT reconstruction pipeline using a self-supervised, coordinate-based learning framework inspired by Neural Radiance Fields (NeRF).
• SpeRF independently trains a multi-layer perceptron (MLP) to estimate skipped SPECT projection views for each scan.
• Tested SpeRF with down-sampling factors (DFs=2, 4, 8) on Lu-177 phantom and clinical SPECT/CT datasets ([177Lu]Lu-DOTATATE and [177Lu]Lu-PSMA-617), comparing reconstructions against 'Full', 'Partial', and 'LinInt' methods.

Main Results:

• SpeRF projections showed lower Normalized Root Mean Squared Difference (NRMSD) compared to linear interpolation (LinInt) projections across phantom and patient studies.
• At DF=4, SpeRF reconstructions outperformed LinInt and Partial methods in Contrast-to-Noise Ratio (CNR) for lesions and organs in both DOTATATE and PSMA-617 studies.
• SpeRF demonstrated good quantitative accuracy, with count recovery close to 'Full' acquisition, outperforming LinInt, especially for organs.

Conclusions:

• SpeRF enables significant reduction in SPECT acquisition time (up to 4x) while preserving quantitative accuracy in clinical protocols.
• The self-supervised nature of SpeRF eliminates the need for extensive training datasets, processing data independently per patient.
• This method is particularly beneficial for low-count SPECT imaging and protocols requiring multiple bed positions, such as whole-body imaging.