Transfer Learning-based Attenuation Map Generation for Brain SPECT Using Simulation and Clinical Data

Abstract

Current attenuation correction (AC) of brain SPECT (Single-photon emission computed tomography) remains challenging in routine clinical practice due to the potential mismatch between SPECT and CT and the increase of radiation absorbed dose from CT. CT-less AC methods have been reported in SPECT based on deep learning (DL). However, DL-based AC methods usually require a large amount of training data while the collection of clinical data is challenging. Transfer learning (TL) has been introduced to develop robust target models via fine-tuning (FT) strategies, using a small set of target training data. This study aims to demonstrate the feasibility of estimating attenuation maps for SPECT based on a conditional generative adversarial network (cGAN) using a small amount of clinical data and a large amount of simulated data. The network was firstly trained by paired simulated none-attenuation-corrected (NAC) SPECT and attenuation map. Then the pre-trained network was fine-tuned by paired clinical NAC SPECT and attenuation map (TLAC). Finally, the challenges, opportunities, and barriers of TLAC were evaluated and discussed.