LOW DOSE CT ENHANCEMENT USING DEEP LEARNING METHOD
Deep learning has been widely applied in traditional medical imaging tasks like segmentation and registration. Some fundamental CNN based deep learning methods have shown great potential in low dose CT (LDCT) enhancement. This study applied U-Net++ model to enhance CT images with low dose and compared the performance of U-net++ and U-net quantitatively and qualitatively.
30 patient CT images were chosen as the ground truth in the training process. Under-sampled projections were simulated from the ground truth volumes with an uniform distribution. LDCT was then reconstructed from the under-sampled projections using the ASD-POCS TV algorithm with 40 iterations and was treated as the input of the models. The U-net++ model was improved based on U-net model by connecting the decoders, reserving better dense feature along skip connections. Deep supervision (DS) were used to make a combined loss between each upper node and the ground truth to enhance the image feature preserving capacity. U-net was used as standard model for comparison. L1 loss and structure similarity (SSIM) loss were used in different attempts. The generated images were compared quantitatively using SSIM and peak signal-to-noise ratio (PSNR).
Both models succeeded to improve the quality of the low dose CT images. The U-net++ model trained with MSE loss had best average PSNR of 17.8 on the test dataset and average SSIM of 0.779 in terms of the whole images compared with the original under-sampled LDCT with SSIM of 0.532 and PSNR of 16.7. U-net model trained using L1 loss had the best average SSIM of 0.756 and the average PSNR of 17.5. Conclusion: Deep learning method showed its potential dealing with the high dose caused by modern CT technique. Different CNN model could influence the quality of the generated image on different evaluation criterions.
This work is licensed under a Creative Commons Attribution-Noncommercial-No Derivative Works 3.0 United States License.
Rights for Collection: Masters Theses
Works are deposited here by their authors, and represent their research and opinions, not that of Duke University. Some materials and descriptions may include offensive content. More info