Diversified Training Manifolds and Augmented Testing for Improved Deep Learned Radionuclide Reconstruction

Deep neural network feed-forward architectures have recently begun exploratory adoption for image reconstruction in radionuclide imaging. However, there is a lack of analysis on the impact of the training data in terms of object variability and count level on the quality of the reconstructions. In this study, the effects of diversifying the training data manifold and test-time augmentation (TTA) are explored for a convolutional encoder-decoder (CED) network intended for the reconstruction of radiotracer distributions using sensor array outputs (SAOs) from a CMOS intraoperative probe. We demonstrate overfitting occurs when a single object type is present in the training data and including more objects in the training domain improved generalisation, yielding better quality reconstructions. The CED network was unable to generalise reliably when tested with count levels not present in the training set, therefore the network must be retrained with desired count levels or else should include estimation of epistemic uncertainty to avoid misleading outcomes. We show that TTA can consistently provide better reconstructions due to the merging of multiple reconstructions after dis-augmentation and match the nRMSE values with higher amounts of training data, serving as a method to reduce the amount of training data required.