PNEUMONIA PREDICTION WITHOUT SEVERITY LABELS: ANALYSIS OF VISION LEARNING INTERDEPENDENCIES

Pneumonia severity assessment has utilized full access to radiological reports containing severity labels. We aimed to study the extent to which severity information can be extracted from metadata (patient information and chest X-rays) without severity labels and investigate the correlation of severity with pneumonia presence binary classification and mask prediction (predicting the area of infection of lungs). We made an ordinal severity scale from 0 to 5 (5 being the most severe pneumonia) with each score corresponding to a range of lung size percentages. We implemented a new methodology of two modified U-net models (chained and parallel), creating severity score predictions based on ground truth and predicted mask size using our ordinal severity scale from 0 to 5. Only the chained model output utilizes probabilistic output from mask predictions and binary classification. We predicted that high-severity infection predictions would be more interrelated to classification and mask output. We found that the chained model (more interrelated by architecture) showed better average severity accuracy than the parallel structure model by 55.5% for severity scores 4 and 5, while the chained and parallel model showed classification validation accuracies of 0.94 and 0.95 respectively. For severity assessment, the chained model had 0.74 average weighted F1-score across all 5 scores, while the parallel model achieved 0.69 average weighted F1-score for severity. The study shows promise of our method, which could be used for future severity classification models to find severity, regardless of severity input predisposition. Future studies should focus on severity accuracy improvement.