TABLE 1. Whole Image Classification Results for Four Downscale and/or Crop Approaches. The Validation Cohort of Images (N = 63 Active EoE; N = 63 non-EoE) Was the Same for Each of the Classifiers. True Positive Rate (TPR; Number of Images Classified as Active EoE / Number of Active EoE Images X 100), True Negative Rate (TNR; Number of Images Classified as non-EoE / Number of non-EoE Images X 100), Accuracy (Number of Images Accurately Classified as Either Active EoE or non-EoE / Total Number of Images X 100), and Predicted Prevalence (Total Number of Images Classified as Active [i.e., True Positive + False Positive Number of Images] / Total Number of Images) for Each Method are Shown. DCNN, Deep Convolutional Neural Network. ACC, Accuracy.
| WHOLE IMAGE PREDICTION | |||||
|---|---|---|---|---|---|
| Original Image | Final DCNN input image size | Active EoE (TPR) | Non-EoE (TNR) | ACC | Predicted Prevalence (PP) |
| Full Image | 1000x1000 (Downscale) | 74.6% | 96.8% | 85.7% | 0.39 |
| Full Image | 224x224 (Downscale) | 65.1% | 88.9% | 77.0% | 0.38 |
| Patch = 448x448 | 224x224 (Downscale) | 82.5% | 87.3% | 84.9% | 0.48 |
| Patch = 224x224 | 224x224 | 82.5% | 77.8% | 80.2% | 0.52 |