Collège St-Michel

Abstract Background / AimsPatient motion during positron emission tomography (PET) imaging can corrupt the image by causing blurring and quantitation error due to misalignment with the attenuation correction image. Data-driven techniques for tracking motion in PET imaging allow for retrospective motion correction, where motion may not have been prospectively anticipated.MethodsA two minute PET acquisition of a Hoffman phantom was acquired on a Bi- ograph mCT Flow, during which the phantom was rocked, simulating periodic motion with varying frequency. Motion was tracked using the sensitivity method, the axial centre-of-mass (COM) method, a novel 3D-COM method, and the principal component analysis (PCA) method. A separate two minute acquisition was acquired with no motion as a gold standard. The tracking signal was discretised into 10 gates using k-means clustering. Motion was modelled and corrected using the reconstruct-transform-add (RTA) technique, leveraging Multimodal Image Registration using Block-matching and Robust Regression (Mirorr) for rigid registration of non- attenuation-corrected 4D PET and Software for Tomographic Image Reconstruction (STIR) for PET reconstructions. Evaluation was performed by segmenting white matter (WM) and grey matter (GM) in the attenuation correction computed tomography (CT). The mean uptake in the region of GM was compared with that in the WM region. Additionally, the difference between the intensity distributions of WM and GM regions was measured with the t-statistic from a Welch's t-test.ResultsDifference in the mean distribution of WM to GM ranked the techniques in order of efficacy: no correction, sensitivity, axial-COM, 3D-COM, PCA, no motion. PCA correction had a great WM/GM separation measured by the t-value than the no motion scan. This was attributed to interpolation blurring during motion correction reducing class variance.ConclusionOf the techniques examined, PCA was found to be most effective for tracking rigid motion. The sensitivity and axial-COM techniques are mostly sensitive to axial motion, and so were ineffective in this phantom experiment. 3D-COM demonstrates improved transaxial motion sensitivity, but not to the level of effectiveness of PCA.

A significant proportion of patients scanned in a clinical setting have follow-up scans. We show in this work that such longitudinal scans alone can be used as a form of 'free' self-supervision for training a deep network. We demonstrate this self-supervised learning for the case of T2-weighted sagittal lumbar Magnetic Resonance Images (MRIs). A Siamese convolutional neural network (CNN) is trained using two losses: (i) a contrastive loss on whether the scan is of the same person (i.e. longitudinal) or not, together with (ii) a classification loss on predicting the level of vertebral bodies. The performance of this pre-trained network is then assessed on a grading classification task. We experiment on a dataset of 1016 subjects, 423 possessing follow-up scans, with the end goal of learning the disc degeneration radiological gradings attached to the intervertebral discs. We show that the performance of the pre-trained CNN on the supervised classification task is (i) superior to that of a network trained from scratch; and (ii) requires far fewer annotated training samples to reach an equivalent performance to that of the network trained from scratch.