ارتقاء كيفيت تصاويرPET-CT با دوز پايين با استفاده از روش يادگيري عميق

In recent decades, PET-CT imaging has played a vital role in the diagnosis an‎d treatment of diseases as one of the most advanced molecular imaging techniques. However, the quality of PET-CT images at low radiation doses is limited due to noise an‎d quality degradation, which particularly poses challenges in the accurate diagnosis of diseases an‎d medical treatments, especially in sensitive groups such as children. Improving the quality of low-dose PET-CT images using deep learning techniques, particularly convolutional neural networks an‎d generative adversarial models, has been proposed as an innovative solution. In previous studies, deep learning models such as 3D U-Net have been used to improve the quality of low-dose PET-CT images. These methods have shown better performance compared to single-input models by reducing noise an‎d improving metrics such as PSNR an‎d RMSE. Specifically, multi-channel models have been able to reduce RMSE by 22.22% an‎d significantly increase PSNR, bringing the image quality to the level of stan‎dard-dose images. However, challenges such as reliance on diverse data, higher computational complexity compared to single-input models, an‎d limited generalizability to other devices an‎d clinical conditions remain issues.This study aims to improve the quality of low-dose PET-CT images by using the deep learning model ResUNet 3D. The data used in this study consists of 219 images from the head an‎d neck region at three different dose levels (2%, 5%, an‎d 10%), with preprocessing steps such as normalization, resizing, an‎d noise removal applied. The proposed model was designed by combining the ResNet an‎d U-Net architectures an‎d eva‎luated based on the metrics RMSE, PSNR, an‎d SSIM. The results showed that the best performance was observed at epoch 35, where RMSE improved to 0.0069 ± 0.0025 for the 10% dose, noise was reduced, an‎d the structural details of the images were preserved. Furthermore, PSNR increased from 22 to 43, indicating a significant improvement in the quality of the reconstructed images. These results demonstrate the high potential of the proposed model in reducing noise, improving reconstruction accuracy, an‎d enhancing the quality of low-dose medical images.The findings of this study indicate that the ResUNet 3D model outperformed similar models such as HighResNet in terms of PSNR, SSIM, an‎d RMSE. Specifically, the proposed model was able to reduce noise an‎d increase image reconstruction accuracy. The reasons for these improvements include more precise processing of spatial an‎d contextual features of low-dose images an‎d the use of residual blocks an‎d larger training datasets. The advantages of this model include noise reduction, improved reconstruction accuracy, an‎d its potential application in reducing radiation dose without compromising image quality in clinical settings. Despite limitations such as the need for large training datasets an‎d powerful hardware resources, the ResUNet 3D model holds significant potential in optimizing the quality of medical images an‎d reducing radiation dose.

5

حسين عربي