University of Twente Student Theses


Invertible recurrent inference machines for low-dose computed tomography

Gurusamy Muthuvelrabindran, A.M. (2021) Invertible recurrent inference machines for low-dose computed tomography.

[img] PDF
Abstract:Computed tomography (CT) is one of the most essential tools in medical imaging. The carcinogenic properties of x-rays motivate the development of techniques that reduce radiation exposure. Low-dose CT is one such approach but due to reduced radiation dose, it produces noisy projection data which in turn lowers the quality of the reconstructions. We turn to deep learning approaches to build quality reconstructions from noisy low-dose CT data. In this work, we investigate the use of invertible recurrent inference machines (iRIM) for low dose CT. Three iRIM models with likelihood gradient definitions of varying complexities were designed and trained uniformly on the LoDoPaB dataset. The image gradient iRIM model had the least complex likelihood gradient definition followed by the adjoint and FBP gradient iRIM models. Our adjoint gradient iRIM model performed the best and obtained an SSIM and PSNR mean and standard deviation of 0.8541 ± 0.1394 and 35.93 ± 4.74 dB on the LoDoPaB test data. It also achieved the best SSIM average of 0.8692 and secured an overall 4th position in the LoDoPaB challenge. Furthermore, the generalization capabilities of the developed iRIM models were tested on three chosen categories – anatomy, low-dose simulation noise level and x-ray source beam geometry. Amongst the iRIM models, the FBP gradient iRIM model proved to be the most capable on this front. The iRIM models were able to produce good results on the generalization capability tests but the performance degraded when the model had to handle high level noise contaminations. In conclusion, the iRIM framework proved to be suitable for low-dose CT but there are still a few scopes of improvement that could be considered to further enhance its robustness.
Item Type:Essay (Master)
UMC, Utrecht, Netherlands
Faculty:EEMCS: Electrical Engineering, Mathematics and Computer Science
Subject:54 computer science
Programme:Electrical Engineering MSc (60353)
Link to this item:
Export this item as:BibTeX
HTML Citation
Reference Manager


Repository Staff Only: item control page