Accessible, Affordable and Low-Risk Lungs Health Monitoring in Covid-19: Deep Cascade Reconstruction from Degraded LR-ULDCT

We present deep cascade reconstruction of degraded low-resolution ultra-low-dose computed tomography (LR-ULDCT) chest images to restored and super-resolved (SR) ULDCT as accessible, affordable, and relatively less hazardous recourse for lungs health monitoring in COVID-19; when compared to relatively less available, costly, and high radiation dose high-resolution CT (HRCT). The degraded LR-ULDCT is first restored with unsupervised dictionary-based deep residual learning network that handles degradations along with Poisson noise found in CT data. The restored version is given to SR network that increases its spatial resolution by minimizing adversarial loss between LR-ULDCT and reconstructed SR-ULDCT within minimax game. It is then fed for segmentation which is achieved by additional block of convolution, Leaky-ReLU, and batch-normalization in U-Net. Thus restored segmented SR-ULDCT estimates presence of ground glass opacity and facilitates monitoring of lungs health at par HRCT. Comparative experiments and ablation study are presented using synthetic and real COVID-19 data.