Noise Reduction of Scintillation Camera Images Using UNET: A Monte Carlo Simulation Approach

Sammanfattning: Background The project aims to reduce the noise in planar 111-In projections with a machine learning model. Method Sixteen base phantoms were used to create 696 phantoms in XCAT. An anterior planar projection was simulated with 111-In and ten uptake curves for each phantom in SIMIND. From the 6960 simulated planar projections 55600 sample and label pairs were created by adding Poisson distributed noise, scaling, rotation, and shift. The label was the noiseless representation of the sample. The normalized sample and label pairs were used to train four UNETS with different hyperparameters. The UNETs were evaluated with the metrics mean NRMSE (normalized root mean square error), mean PSNR (peak signal to noise ratio) and mean MSSIM (mean structural similarity index). A fifth UNET was trained with 55600 sample and label pairs without normalization. A UNET trained with normalized and un-normalized training data was compared to a standard Butterworth filter. The comparison was performed using the normalized profiles from the different images. Results The smallest mean PSNR was 53. The largest mean NRMSE was 0.19 and mean MSSIM was 0.81. The difference in the means between the UNETs for the different evaluation metrics was evaluated with a one-way ANOVA test for the test data containing no augmentation added except scaling. The one-way ANOVA test showed no statistical difference between the means for the different UNETs with regards to the calculated PSNR and NRMSE values. However, a statistically significant difference was found between the means of the different UNETs for the calculated MSSIM values. The Butterworth profile was more consistent with the ground truth label for the test data than the profiles from the UNETs trained with normalized and un-normalized data. The UNET from the un-normalized resulted in blurrier predictions compared to the UNET trained with normalized data. Conclusions The UNET can reduce the noise in scintillation camera images with a high noise level. However, the UNET trained could not recover details such as the ribs and was overfitted since the size of the spleen in the predictions was not consistent with the labels. Further work is needed to optimize the training data and the architecture to recover details lost in the imaging processing and minimize overfitting. The images from the UNET trained with normalized training data were not as blurry as the images filtered with the Butterworth filter.