CNN-Based Denoising of Industrial CT Images: A Residual-Learning Benchmark Against Classical Filters on Synthetic Phantoms

Abstract X-ray computed tomography (CT) lets engineers look inside manufactured parts without cutting them open. The catch is that the images are noisy, and the noise gets worse as the X-ray dose drops. That grain can hide the small pores and inclusions an inspector is actually looking for, and classical filters, which follow fixed rules, tend to blur the very edges they should protect. In this project, carried out with co-author Professor Mayank Goswami for course PHC-351 (Project 9, Category B), I built a fully synthetic, non-biological benchmark of 5,000 industrial phantoms (256 × 256 pixels). The phantoms have five outer shapes, bright rims, 0 to 6 dense inclusions and 0 to 9 pores, and the noise combines Poisson photon statistics with Gaussian electronic noise at three dose levels. I trained a 17-layer residual convolutional network (DnCNN, 556,096 parameters) on 4,200 images, validated it on 400, and compared it with Gaussian, median and non-local-means filters on 400 images it had never seen. DnCNN reached a mean peak signal-to-noise ratio (PSNR) of 50.28 dB and a structural similarity (SSIM) of 0.998. For comparison, non-local means scored 45.56 dB and 0.910, the median filter 41.38 dB and 0.964, the Gaussian filter 37.44 dB and 0.857, and the noisy input 37.93 dB and 0.809. DnCNN beat the best classical filter on all 400 test images (smallest margin + 1.87 dB, mean + 4.72 dB) and at every dose level (low 48.92 dB, medium 50.06 dB, high 50.88 dB). One caveat matters: training and test noise come from the same simulator, so these numbers are a best case, and testing on real scanner data is the obvious next step.