Generating PET Attenuation Maps via Sim2Real Deep Learning–Based Tissue Composition Estimation Combined with MLACF

• DOI: 10.1007/s10278-023-00902-0
• Publication Date: 2024-01-10T18:01:32Z
• Authors: Tetsuya Kobayashi; Yui Shigeki; Yoshiyuki Yamakawa; Yumi Tsutsumida; Tetsuro Mizuta; Kohei Hanaoka; Shota Watanabe; Daisuke Morimoto‑...; Takahiro Yamada; Hayato Kaida; Kazunari Ishii
• Keywords: Correction for attenuation

Deep learning (DL) has recently attracted attention for data processing in positron emission tomography (PET). Attenuation correction (AC) without computed (CT) is one of the interests. Here, we present, to our knowledge, first attempt generate an attenuation map human head via Sim2Real DL-based tissue composition estimation from model training using only simulated PET dataset. The DL accepts a two-dimensional non-attenuation-corrected image as input and outputs four-channel tissue-composition soft tissue, bone, cavity, background. Then, generated by linear combination maps and, finally, used scatter+random initial estimate reconstruction maximum likelihood factor (MLACF), i.e., refined MLACF. Preliminary results clinical brain showed that proposed tended anatomical details inaccurately, especially neck-side slices. However, it succeeded estimating overall structures, quantitative accuracy with AC was comparable CT-based AC. Thus, approach combined MLACF also promising CT-less approach.