Using Self-Supervised Learning To Improve Deep Learning-Based Analysis of 4D PET Imaging

Abstract

Dynamic positron emission tomography (dPET) imaging requires an accurate arterial input function (AIF) for quantitative analysis. However, traditional methods for AIF measurement require invasive blood sampling. Recent deep learning models can predict the AIF from the 4D PET data itself, eliminating the need for invasive blood sampling. However, their performance often degrades with limited amount of training data or changes in acquisition protocols. To address this, we introduce a two-stage self-supervised learning (SSL) approach. First, a convolutional encoder that is pretrained on unlabeled dPET data using an autoencoder. Then the pretrained weights are used to initialize the encoder which is fine-tuned with supervised learning to map the dPET data to the AIF. We evaluate standard, denoising and masked autoencoder pretext tasks on a preclinical mouse [\textsuperscript{18}F]FDG dPET dataset. Compared to a baseline DLIF model trained from scratch, the SSL pretrained models achieves improved AIF prediction accuracy. For example, the denoising autoencoder pretraining raised the $R^2$ from about 0.91 to 0.93 and reduced the mean squared error by roughly 35$\%$. These gains can be transferred into more reliable tracer kinetic parameter estimates without any invasive blood sampling. In summary, using SSL to exploit the abundance of unlabeled PET data greatly enhances non-invasive AIF estimation, thus advancing the robustness and applicability of dynamic PET imaging.