COVID-19 prognosis with artificial intelligence (AI)
AI-based lung image analysis enhances disease severity assessment, reducing ICU overload with standardized admission criteria for Covid-19 patients. Expanded AI research is vital for integrating it into clinical practice and preparing for future pandemics.
Automated Covid-19 segmentation and quantification of lung involvement using deep learning hold promise. However, there are notable disparities between clinicians' and AI communities' studies regarding patient care for Covid-19. Therefore, integrating AI into clinical practice requires addressing challenges in standardized severity classification, lung lesion characterization, multi-modal imaging data integration, robust quantification of long Covid severity, and understanding acute-to-chronic phases. These steps are crucial for optimizing patient care.
This study aimed to develop a modular AI-based approach for modeling a patient's current state and predicting the short and long progression of Covid-19 patients. The specific objectives were establishing a severity assessment system based on the WHO clinical progression scale, including chest X-rays, to predict whether patients need intubation after seven days based on baseline medical images, and to create an AI model to predict the severity of Covid-19 disease in the chronic phase.
Our AI model, AssessNet-19 achieved an F1-score of 0.76±0.02 for severity classification in the evaluation set, which was superior to the three expert thoracic radiologists and the single-class lesion segmentation model. In addition, AssessNet-19 automated multi-class lesion segmentation obtained a mean Dice score of 0.70 for Ground-glass opacity (GGO), 0.68 for consolidation, 0.65 for pleural effusion, and 0.30 for band-like structures compared to ground truth. Moreover, it achieved a high agreement with radiologists for quantifying disease extent with Cohen's Kappa of 0.94, 0.92, and 0.95.
Transferable vision transformers guided by GGO and CON mask achieved an F1 score of 0.6972 and an AUC of 0.7452 for the 7-day intubation prediction task in the test set. It outperforms the vision transformer trained on DRRs and tested on XRs with an F1 score of 0.5819 and an AUC score of 0.6785. Moreover, the transferable vision transformer guided by GGO and CON mask generates natural attention maps along with prediction results, showing the important regions for model prediction.
Specific contribution to tackle the current pandemic
We provided a multi-center Covid-19 dataset: curated, labeled, diverse radiological, clinical and laboratory data, reduces biases, enhances generalizability with varied cases, severities, CT scan sources, and contrast use.
A novel AI multi-class radiomics model including seven lung lesions to assess disease severity based on the WHO-CPS scale more accurately determines the severity of Covid-19 patients than a single-class model and radiologists' assessment.
AI-multi-omics-based Prognostic Stratification of COVID-19 Patients in Acute and Chronic State