The Performance of Deep U-Net Pre-Clinical Organ-wise Segmentation in the Presence of Low Counting Statistics.Download PDF

25 Jan 2020 (modified: 05 May 2023)Submitted to MIDL 2020Readers: Everyone
Abstract: Micro-PET-CT allows non-invasive monitoring of biological processes, disease progression and therapy response. Morphological information provided by the CT allows organ / tissue delineation for subsequent quantification of the physiological information depicted by the PET. Deep learning with convolutional neural networks (CNNs) has achieved state-of-the-art performance for automated medical image segmentation and utilized successfully by our group in Micro-PET-CT (figure 1). The robustness of such approaches in the presence of noise addition / dose reduction of the CT data has not been explored. We thus simulate dose reduction of pre-clinical CT images using a Poisson noise model and evaluate the effect of segmentation performance with increasingly lower dose for 7 regions (skeleton, kidney, bladder, brain, lung, muscle and fat). data in the preclinical model with increase dose reduction. It can be observed that the accuracy of the segmentation measured by the DICE coefficient falls off as we simulate the reduction of CT dose. A 50% dose reduction was observed for all 5 test subjects to result in a mean (across all 7 organs) percentage reduction in DICE <25%. Adequate performance however is still observed in the DICE coefficient with a dose reduction of 30%, only an average of ~10% reduction in DICE is observed. This may have implications for utilising reduced dose CT coupled with a deep CNN for segmentation if the CT component is used to anatomically locate physiology on PET data.
Paper Type: methodological development
TL;DR: Exploring the effectiveness of deep learning U-Net for automated organ-wise pre-clinical image segmentation in the presence of low counting statistics from CT images
Track: short paper
Keywords: deep learning, pre-clinical imaging, PET-CT, image segmentation, artificial intelligence
6 Replies

Loading