Figure 7. Repurposing the NFT dataset for object-detection models yields similar overall performance to the segmentation approach.

A) The precision-recall curve plots the tradeoff between the YOLO model’s object-detection precision and recall across all potential confidence thresholds aggregated across all objects in the test dataset. B) Plot of aggregated object-wise F1 versus confidence threshold; low (permissive) thresholds yielded the highest F1 scores. C) WSI-level NFT counts from the YOLO model correlated with CERAD-like semi-quantitative scoring, similar to the segmentation-derived NFTDetector scores in Figure 5b. **: 0.001 < p <= 0.01 by Welch’s t-test. D) Example tiles from different ROIs overlaid with ground truth (cyan) and predicted (yellow) bounding boxes. The values highlighted in yellow are confidence scores for the predictions. Reference B) to determine which NFTs would be detected.