Denoising Low Dose CT Images Using Enhanced Autoencoder Deep Learning Model: A Phantom Study

Reducing radiation dose in Computed Tomography (CT) inherently leads to increased image noise, which can compromise clinical interpretation. To address this limitation in Low Dose CT images (LDCT), numerous computational techniques particularly convolutional neural networks (CNN) have been explored for CT image enhancement. Various CNN architectures with different convolutional layer designs have been developed to improve image quality. This study evaluates the effectiveness of a developed CNN-based autoencoder for denoising LDCT images while preserving diagnostic integrity. An end-to-end CNN autoencoder with information bottleneck (IB) pipeline was designed and trained using CT phantom images with varying noise levels, aiming to suppress noise while maintaining anatomical detail. The model's denoising performance was quantitatively assessed using metrics such as Peak Signal-to-Noise Ratio (PSNR), Structural Similarity Index Measure (SSIM), Root Mean Squared Error (RMSE), and Mean Absolute Error (MAE). The CNN autoencoder was compared against previous other deep learning-based denoising models, including U-Net and the Enhanced Deep Convolutional Neural Network (EDCNN). Results shows that there is a significant improvement in LDCT image quality, with PSNR increasing from 16 dB in noisy inputs to 30.12–34.83 dB post-denoising. SSIM scores also improved, indicating enhanced structural preservation. While the proposed model outperformed U-Net and EDCNN in terms of noise reduction and structural fidelity, its performance was slightly below that of the Classification Densely Connected Residual Network. Overall, the findings demonstrate that CNN-BI based autoencoders offer an effective solution for LDCT denoising, supporting dose reduction strategies without compromising diagnostic accuracy, despite their relatively high computational demand.