Deep learning-based CT lesion analysis and radiological burden stratification in clinically diagnosed ventilator-associated pneumonia.
Authors
Affiliations (1)
Affiliations (1)
- School of Medical Sciences, Universiti Sains Malaysia, Health Campus, Kubang Kerian, Kelantan, Malaysia.
Abstract
Ventilator associated pneumonia (VAP) is a major complication in the intensive care unit (ICU) and requires objective computed tomography (CT) evaluation. Therefore, this article conducted a retrospective research analysis and developed a multimodal deep learning framework. Firstly, this retrospective study included 512 patients clinically diagnosed with VAP, and then utilized a developed multimodal deep learning framework for voxel-level lesion segmentation, multi-label radiological discovery classification, and radiological burden grading. This framework integrates the anatomical priors of TotalSegmenter, a three-dimensional (3D) Swin UNETR multitasking network, and cross attention fusion of CT features and ventilation parameters. Ultimately, the study showed that the Dice score for internal validation was 0.847, the macroscopic mean area (AUC) under the subject's working characteristic curve was 0.919, and the weighted kappa was 0.826. External validation showed a Dice score of 0.821, with no significant difference between scanner manufacturers. After adjusting for clinical covariates, including acute physiology and chronic health assessment II (APACHE II) scores, there was still a significant correlation between radiation load level and 28 day mortality (hazard ratio [HR] = 2.34, <i>P</i> < 0.001), supporting its potential prognostic relevance.