Denoising of Chest CT (Computed Tomography) images with a U-Net based Convolutional Autoencoder

In this master’s thesis work, a Deep Learning-based denoising algorithm was developed to address the problem of dose reduction in chest CT scans and the resulting increase in noise. In particular, a U-Net based Convolutional Autoencoder was implemented. To overcome the problem of data scarcity (common in AI applications in the medical field) a two-stage training strategy was used. For the first stage, images of the commercial Catphan phantom were acquired, varying the acquisition parameters to obtain the low-dose and high-dose image pairs used for the supervised training of the model. Following this first phase, a transfer learning strategy was implemented to bridge the gap between the phantom and the clinical image domains. The analyses performed on chest LDCTs, before and after the application of the algorithm, demonstrated a reduction in the noise magnitude by a factor of (3.4 +/- 0.6). To further evaluate the capabilities of the model to improve perceived image quality without impairing the visibility of nodules and structures inside the lung, the same images were scored by two experienced radiologists. The scores obtained confirmed the reduction in noise, which resulted in an increase in perceived image quality. Altogether, the results obtained in this thesis demonstrated the possible utility of this method in a lung cancer screening context.