Deep Learning‑based Ultra-high-resolution CT imaging of Viral Pneumonia at Admission and after Discharge

Background Advancement in deep learning has introduced significant potential for enhancing CT image quality without increasing patient radiation exposure. In this study, we sought to compare deep learning‑based ultra-high-resolution CT (UHRCT-DL) findings of viral pneumonia at admission and after discharge with that of HRCT images. Methods A total of 51 inpatients (mean age 66.78 years; 33 males) of viral pneumonia underwent 102 CT scans at admission and after discharge. A deep learning-based super-resolution model, incorporating a dual-branch architecture for super-resolution and gradient guidance, was used to generate UHRCT-DL. UHRCT-DL and HRCT images were systematically reviewed for viral pneumonia CT findings, including ground-glass opacity (GGO), reticulation, tree-in-bud opacities, consolidation, linear bands, bronchiectasis, and bronchiectasis. Subjective and objective CT image quality was evaluated using a five-point Likert scale (− 2 to 2) and lung signal-to-noise ratios (SNRs). Results The score of clarity of CT findings was significantly higher on UHRCT-DL for all CT findings at admission and after discharge. Compared with HRCT as reference image, the most frequently observed additional/different CT findings on UHRCT-DL at admission were crazy paving pattern (14/51, 27%) and tree-in-bud opacities (8/31, 26%), whereas reticulations (15/51, 29%) and bronchiolectasis (12/44, 27%) were most observed additional/different CT findings after discharge. The subjective and objective image quality of UHRCT-DL was superior to that of HRCT. UHRCT-DL algorithm significantly lowered the level of image noise and improved SNR (19.96 ± 6.46 vs 41.35 ± 11.49, p  < 0.001). Conclusions The deep learning‑based UHRCT provided a more precise depiction of CT features of viral pneumonia, that better reflects the inflammatory changes during acute phase and early fibrotic changes during recovery.