Automated Deep Identification of Radiopharmaceutical Type and Body Region From Pet Images Publisher



Ghafari A^{1, 2} ; Sheikhzadeh P^{1, 3} ; Seyyedi N⁴ ; Abbasi M³ ; Ahamed S^{5, 6} ; Ay MR^{1, 7} ; Rahmim A^{5, 6}
Source: Iranian Journal of Nuclear Medicine Published:2024

Abstract

Introduction: A deep learning pipeline consisting of two deep convolutional neural networks (DeepCNN) was developed, and its capability to differentiate uptake patterns of different radiopharmaceuticals and to further categorize PET images based on the body regions was explored. Methods: We trained two sets of DeepCNN to determine (i) the type of radiopharmaceutical ([18F]FDG and [68Ga]Ga-PSMA) used in imaging (i.e., a binary classification task), and (ii) body region including head and neck, thorax, abdomen, and pelvis (i.e., a 4-class classification task), using the 2D axial slices of PET images. The models were trained and tested for five different scan durations, thus studying different noise levels. Results: The accuracy of the binary classification models developed for different scan duration levels was 98.9%–99.6%, and for the 4-class classification models in the range of 98.3%–99.9 ([18F]FDG) and 97.8%–99.6% ([68Ga]Ga-PSMA). Conclusion: We were able to reliably detect the type of radiopharmaceutical used in PET imaging and the body region of the PET images at different scan duration levels. These deep learning (DL) models can be used together as a preliminary input pipeline for the use of models specific to a type of radiopharmaceutical or body region for different applications and for extracting appropriate data from unclassified images. Copyright © 2024 The Authors.