Effective Noise Reduction in Pediatric CT Data: Application of Fine-Tuning Techniques Utilizing Adult CT Data

High noise in low-dose pediatric CT impairs the performance of deep learning denoising models. This study investigates optimizing denoising by fine-tuning a denosing convolution neural network (DnCNN), pre-trained on adult CT data, with small, age-stratified pediatric datasets. We sequentially fine-tuned the final 1, 3, 5, and 7 convolutional layers and evaluated performance using the peak signal-to-noise ratio (PSNR) and the structural similarity index measure (SSIM). Results showed that fine-tuning significantly improved denoising for all age groups.(p<0.05) Notably, performance in younger age groups improved as more layers were fine-tuned, whereas for adolescents, fine-tuning a single layer was optimal to avoid overfitting. Our findings demonstrate that transfer learning is an effective approach and highlight that an age-specific fine-tuning depth is crucial for maximizing denoising performance in pediatric CT.