Purpose: Although deep learning (DL) methods have shown promising performance in low-dose CT denoising, concerns have been raised regarding image fidelity loss. To better preserve the image fidelity while maintaining the strong statistical learning ability of DL, we proposed a hybrid method to combine DL with regularized optimization and then tested its performance on low-dose lung CT images.
Methods: Lung CT images of 42 patients were reconstructed using filtered back-projection from full-dose and corresponding low-dose projections, which were downloaded from TCIA. The patients were split into train, test, and validation sets comprising 34, 4, and 4 patients, respectively. A denoising U-Net was trained to predict full-dose images from low-dose images. Low-dose images in test sets were fed into the trained U-Net to produce baseline DL-denoised images. Recovered images using the hybrid method were obtained by minimizing an objective function that penalized the L2 distance with denoised images and the L2 discrepancy with low-dose projections. To objectively evaluate the reconstruction quality, we trained an airway segmentation network on a separate lung CT dataset. Automated airway segmentation was then performed on the full-dose, DL-denoised, and hybrid recovered images using the trained model. Using the airway segmentation on the full dose images as the ground truth, we calculated the Dice coefficient and detectability of the DL-denoised and hybrid recovered image airway masks.
Results: On average, hybrid recovered images showed an improvement of 0.00435 in Dice coefficient and 0.0165 in detectability over DL-denoised images, indicating the superior performance of our proposed method over the naive DL method.
Conclusion: This work demonstrated a novel hybrid-denoising method combining the classical regularized reconstruction method with DL for low-dose CT reconstruction. Using automated airway segmentation for objective assessment, we showed that the hybrid method is superior to DL alone in preserving the imaging details while suppressing the noise.