TY - GEN
T1 - Federated Learning-based Deep Learning Model for PET Attenuation and Scatter Correction
T2 - 2021 IEEE Nuclear Science Symposium and Medical Imaging Conference, NSS/MIC 2021
AU - Shiri, Isaac
AU - Sadr, Alireza Vafaei
AU - Sanaat, Amirhossein
AU - Ferdowsi, Sohrab
AU - Arabi, Hossein
AU - Zaidi, Habib
N1 - Publisher Copyright:
© 2021 IEEE.
PY - 2021
Y1 - 2021
N2 - Scatter compensation (SC) and attenuation correction (AC) are vital steps towards quantitative PET image analysis. Recently deep learning algorithms were applied in the image domain for SC and AC. To build a generalizable and reproducible deep learning model a large dataset is needed to tune millions of model parameters. Because of the sensitivity of medical images and strict regulations, gathering a large dataset for deep learning model training is the main challenge in medical applications. In this work, non-attenuation corrected, and CT-based attenuation corrected 18F-FDG PET images of 300 patients were enrolled. Data consisted of 50 patients from 6 different centers which scanners, image acquisition, and reconstruction vary across the different centers. We used a deep residual network as the main architecture. The network consists of twenty convolutional layers in which feature extraction (low, medium, and high) was performed by dilation kernel. For model evaluation, voxel-wise mean error (ME), mean absolute error (MAE), relative error (RE%), absolute relative error (ARE%) and structural similarity index (SSIM) were calculated between ground truth CT-based attenuation/scatter corrected and predicted PET images. We implemented server aggregate federated learning workflow, which included 3 steps: (1) central global model distributed through different departments, then (2) models will be trained in each center separately and finally (3) local trained models returned to central server and model aggregated as central global models. Steps 1-3 are repeated until the model is fully trained and converged. Quantitative analysis showed ME of 0.05±0.1, MAE of 0.43±0.01, RE of 2.74±5.7%, ARE of 15.0±8.8% and SSIM of 0. 90±0.09 in the test set. In this study, we built a deep learning-based AC/SC model for PET images using data emanating from 6 different centers without sharing the datasets. Federated learning algorithms provide the opportunity to build a model using multicenter datasets without sharing data.
AB - Scatter compensation (SC) and attenuation correction (AC) are vital steps towards quantitative PET image analysis. Recently deep learning algorithms were applied in the image domain for SC and AC. To build a generalizable and reproducible deep learning model a large dataset is needed to tune millions of model parameters. Because of the sensitivity of medical images and strict regulations, gathering a large dataset for deep learning model training is the main challenge in medical applications. In this work, non-attenuation corrected, and CT-based attenuation corrected 18F-FDG PET images of 300 patients were enrolled. Data consisted of 50 patients from 6 different centers which scanners, image acquisition, and reconstruction vary across the different centers. We used a deep residual network as the main architecture. The network consists of twenty convolutional layers in which feature extraction (low, medium, and high) was performed by dilation kernel. For model evaluation, voxel-wise mean error (ME), mean absolute error (MAE), relative error (RE%), absolute relative error (ARE%) and structural similarity index (SSIM) were calculated between ground truth CT-based attenuation/scatter corrected and predicted PET images. We implemented server aggregate federated learning workflow, which included 3 steps: (1) central global model distributed through different departments, then (2) models will be trained in each center separately and finally (3) local trained models returned to central server and model aggregated as central global models. Steps 1-3 are repeated until the model is fully trained and converged. Quantitative analysis showed ME of 0.05±0.1, MAE of 0.43±0.01, RE of 2.74±5.7%, ARE of 15.0±8.8% and SSIM of 0. 90±0.09 in the test set. In this study, we built a deep learning-based AC/SC model for PET images using data emanating from 6 different centers without sharing the datasets. Federated learning algorithms provide the opportunity to build a model using multicenter datasets without sharing data.
UR - https://www.scopus.com/pages/publications/85132920617
UR - https://www.scopus.com/pages/publications/85132920617#tab=citedBy
U2 - 10.1109/NSS/MIC44867.2021.9875813
DO - 10.1109/NSS/MIC44867.2021.9875813
M3 - Conference contribution
AN - SCOPUS:85132920617
T3 - 2021 IEEE Nuclear Science Symposium and Medical Imaging Conference Record, NSS/MIC 2021 and 28th International Symposium on Room-Temperature Semiconductor Detectors, RTSD 2022
BT - 2021 IEEE Nuclear Science Symposium and Medical Imaging Conference Record, NSS/MIC 2021 and 28th International Symposium on Room-Temperature Semiconductor Detectors, RTSD 2022
A2 - Tomita, Hideki
A2 - Nakamura, Tatsuya
PB - Institute of Electrical and Electronics Engineers Inc.
Y2 - 16 October 2021 through 23 October 2021
ER -