MetaMis: A Study of Identifying Missed Labels or Mislabels of Chest Radiographic Images Using Meta Learning

Chest radiographic images of lungs were useful for assessing the compilations, prognostication of lungs conditions, monitor treatment for various diseases like COVID-19, Lung cancer, Pneumonia etc. Manually mislabeling and missed labeling occurs which fails to find the appropriate diseases and patient’s lung conditions. Artificial intelligence helps to investigate various medical reports and find solutions without human interventions. To diagnose the missed labeling and mislabeling of the chest computerized tomography (CCT) reports, AI based Meta learning algorithms were used. In this study, nearly 15 million reports from 14 different organization databases were combined; it covers the years from 1999 to 2023 and looked for any chest radiographic reports that included appendices. The 3469 CCT reports that had addenda removed because of typos, the error during writing reports, losses of some important section were removed by a thoracic radiologist. Additionally missed or error in diagnoses such as pneumothorax, fibrosis, lung nodules, consolidations, fractures in rib etc. were present in the remaining 279 patient reports. The misdiagnosis and missed diagnosis for these CCT were analyzed with various metrics. Specific findings (sensitivity, specificity, accuracy) associated with the AI based Meta learning included nodules (96.32%, 98.13%, 96.42%), pneumothorax (84.91%, 93.41%, 85.89%), pleural effusion (96.87%, 53.92%, 71.09%), consolidation (97.52%, 91.25%, 96.42%), and rib fractures (87.65%, 95.67%, 94.83%) respectively. When compared to various AI algorithms, Meta Learning can accurately identify mislabelled and overlooked findings and reduces the number of inaccuracies that occur when radiographic findings are discovered and labeled.