Cardiff University | Prifysgol Caerdydd ORCA
Online Research @ Cardiff 
WelshClear Cookie - decide language by browser settings

Deep learning pre-clinical medical image segmentation for automated organ-wise delineation of PET

Smith, Rhodri L., Paisey, Stephen J., Evans, N., Florence, V., Fittock, E., Siebzehnrubl, F. and Marshall, Christopher 2018. Deep learning pre-clinical medical image segmentation for automated organ-wise delineation of PET. Presented at: Annual Congress of the European Association of Nuclear Medicine, Barcelona, Spain, 12-16 October 2019.

Full text not available from this repository.

Abstract

Aim/Introduction: Micro-PET-CT allows non-invasive monitoring of biological processes, disease progression and therapy response. Morphological information provided by CT allows organ / tissue delineation for subsequent quantification of physiological information depicted by 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 MicroPET-CT. 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 6 regions (skeleton, kidney, bladder, brain, lung, muscle and fat). Materials and Methods: Negating electronic noise and subject anatomy, variance in x-ray images may be asserted to quantum noise allowing a model for simulated image noise versus dose (mA) to be constructed. With a mono-energetic x-ray source the mean number of photons (N) incident on the detector is N=Noexp(-projections); No represents x-ray intensities. N is modelled as a Poisson process (λ=N) a term for the projection noise is thus log(N/No). Modulating No serves as a variable to control noise levels (~1/No) in projections. A linear relationship between image variance in a central region of an example image and 1/No was found. As image variance ~ 1/mA, systematically reducing No has the effect of simulating a percentage dose reduction in mA. Simulated dose reduction of 5 test cases (50kVp, 300mA, 0.25x0.25x0.25mm) was performed in the sinogram domain allowing reconstruction of noisy simulated CT images. Dose reduction was performed from 10% - 100% in increments of 10%. Test images were segmented using our previously trained model for whole body pre-clinical image segmentation. The Dice coefficient of the six segmented regions for each test case and each dose reduction was assessed. Results: The percentage reduction in DICE from the ground truth serves as a measure of the reduction in performance of the segmentation with increasing noise. A 50% dose reduction was for all 5 test subjects resulted in a mean (across all 6 organs) percentage reduction in DICE <25%. Conclusion: The pattern in reduction in performance of segmentation as dose was reduced was similar for each tissue. This may have implications for utilizing reduced dose CT coupled with a deep CNN for segmentation if the CT component is used to anatomically locate physiology on PET data

Item Type: Conference or Workshop Item (UNSPECIFIED)
Status: Unpublished
Schools: Medicine
Last Modified: 29 Feb 2020 16:25
URI: http://orca-mwe.cf.ac.uk/id/eprint/126771

Actions (repository staff only)

Edit Item Edit Item