Pre-training and Transfer Learning for Training Set Reduction and Improving Automated Assessments of Clinical PET Image Quality

Acquiring simultaneous positron emission tomography - magnetic resonance (PET-MR) images for memory clinic patients using a fraction of the usual injected dose can help reduce radiation dose concerns. However, this compromises the reconstructed image quality, thus determination of clinical usability is beneficial. A major challenge is the paucity of clinical quality readings. We hypothesise that exploiting easily available quantitative information or using established pre-trained networks to predict clinical assessments could reduce the number of clinically assessed datasets required for training convolutional neural networks (CNNs). In this study, a CNN was pre-trained to predict injected dose of patches extracted from six real patient datasets, reconstructed using 100% of the available data and different thinned datasets down to 0.5% of available data. Using transfer learning with five separate patients, the CNN was used to predict three clinically scored quality metrics: global quality rating. pattern recognition and diagnostic confidence, based on a four-point scale (0-3). This was compared to pre-training a VGG16 network with ImageNet at varying pre-training levels. This work shows test performance can be improved via pre-training compared to using no pre-training. Pre-training the last two convolutional blocks and fully connected layer of a VGG16 backbone achieves a 93 ± 3.7% broad agreement (predicted metric within 1 of clinician score) across all metrics, compared to 81 ± 2.6% when using no pre-training. Pre-training via dose inference achieves a maximum Pearson’s correlation coefficient of 0.90 for global quality rating, compared to 0.61 when using no pre-training. Easily accessible quantitative labels or pre-trained networks may be exploited to predict scarce clinical metrics. Future work will include increasing the clinical dataset number and applying the CNN to superior reconstruction algorithms.