"Citra 3D Rotational Angiography rentan terhadap noise akibat penggunaan wide cone beam x-ray dan terbatasnya jumlah citra untuk rekonstruksi, sehingga diperlukannya proses denoising. Penelitian ini membandingkan algoritma residual encoder-decoder convolutional neural network (RED-CNN) dan Non-Local Means Denoising (NLMD) dalam mengurangi noise pada citra 3D rotational angiography (3DRA) phantom anthropomorphic baik pada citra tanpa tambahan noise (original) maupun citra dengan tambahan noise Gaussian dan Poisson. RED-CNN dilatih menggunakan citra CT scan yang ditambahkan noise dan hasil citra dievaluasi menggunakan peak signal to noise ratio (PSNR), signal difference to noise ratio (SDNR), dan structural similarity index measure (SSIM). Pada citra dengan noise Gaussian, NLMD meningkatkan PSNR sebesar 15,39%–30,44% dan RED-CNN sebesar 7,199%–48,10%. SDNR meningkat 489,3%–21,35% (NLMD) dan 129,1%–263,3% (RED-CNN), sedangkan SSIM meningkat 155,9%–22,59% (NLMD) dan 84,66%–181,9% (RED-CNN). Pada noise Poisson, NLMD meningkatkan PSNR 2,389%–31,91% dan RED-CNN 2,178%–6,108%; SDNR meningkat ~300,0% (NLMD) dan 80,52%–137,9% (RED-CNN); sementara SSIM menurun 8,561%–13,04% (NLMD) dan 5,857%–0,6143% (RED-CNN). Untuk citra original, NLMD meningkatkan SDNR sebesar 134,9% namun menurunkan SSIM sebesar 21,43%. RED-CNN menghasilkan peningkatan SDNR sebesar 76,36% (dilatih dengan noise Poisson) dan 43,02% (Gaussian), serta penurunan SSIM sebesar 12,12% dan 15,41%. Dari sisi efisiensi, NLMD memproses 10 slice dalam 22,30 detik tanpa GPU, sedangkan RED-CNN membutuhkan waktu 30 menit dengan GPU. Hasil menunjukkan bahwa RED-CNN unggul dalam peningkatan kualitas citra, sedangkan NLMD lebih efisien secara komputasi.
3D Rotational Angiography (3DRA) images are prone to noise due to the use of wide cone beam X-ray and the limited number of projections fot reconstructions, making denoising is necessary. This study compares the effectiveness of the Residual Encoder-Decoder Convolutional Neural Network (RED-CNN) and Non-Local Means Denoising (NLMD) in reducing noise in anthropomorphic phantom 3DRA images, both in original images and those with added Gaussian and Poisson noise. The RED-CNN model was trained using CT images with added noise, and evaluation was conducted using Peak Signal to Noise Ratio (PSNR), Signal Difference to Noise Ratio (SDNR) and Structrural Similarity Index Measure (SSIM) metrics. For images with Gaussian noise, NLMD improved PSNR by 15.39%–30.44% and RED-CNN by 7.199%–48.10%. SDNR increased by 489.3%–21.35% (NLMD) and 129,1%–263,3% (RED-CNN), while SSIM increased by 155.9%–22.59% (NLMD) and 84.66%–181.9% (RED-CNN). For Poisson noise, NLMD improved PSNR by 2.389%–31.91% and RED-CNN by 2.178%–6.108%; SDNR increased by approximately 300.0% (NLMD) and 80.52%–137.9% (RED-CNN); while SSIM decreased by 8.561%–13.04% (NLMD) and 5.857%–0.6143% (RED-CNN). For original (noise-free) images, NLMD improved SDNR by 134.9% but reduced SSIM by 21.43%. RED-CNN improved SDNR by 76.36% (trained with Poisson noise) and 43.02% (Gaussian noise), but reduced SSIM by 12,12% and 15.41%, respectively. In terms of efficiency, NLMD processed 10 slices in 22.30 seconds without requiring a GPU, while RED-CNN required 30 minutes and a dedicated GPU. Overall, RED-CNN demonstrated superior image quality enhancement, while NLMD was more computationally efficient."
Depok: Fakultas Matematika dan Ilmu Pengetahuan Alam Universitas Indonesia, 2025