Confidence Score

A post-processing algorithm that eliminates redundant, overlapping bounding boxes for the same object, retaining only the detection with the highest confidence score. In medical imaging, NMS prevents a single lesion from generating multiple false alarms. The algorithm sorts all detections by confidence, selects the highest-scoring box, and suppresses any other box with an Intersection over Union (IoU) exceeding a set threshold (e.g., 0.5). Soft-NMS variants decay scores rather than zeroing them, improving recall for closely clustered abnormalities like micro-calcifications.

A loss function designed to address extreme class imbalance in dense object detection by down-weighting the loss assigned to well-classified, easy examples. Standard cross-entropy loss is overwhelmed by the vast number of easy background regions in radiological scans. Focal loss adds a modulating factor (1 - p_t)^γ to the cross-entropy loss, where γ is a focusing parameter. When a model predicts a background region with high confidence (p_t ≈ 0.9), the loss contribution is scaled down by an order of magnitude, forcing the optimizer to focus on hard, misclassified examples like small nodules.

The fundamental visualization of a detection model's performance across all possible confidence score thresholds. Precision measures the fraction of positive detections that are correct, while recall measures the fraction of all ground truth abnormalities that were found. By sweeping the confidence threshold from 0 to 1, a curve is generated. A model with high area under this curve maintains high precision even at high recall. In clinical CADe systems, the operating point is often chosen to favor high recall (sensitivity) at the expense of some false positives, as missed findings are more dangerous than false alarms.

The Free-Response Receiver Operating Characteristic curve is the standard evaluation metric for lesion detection tasks in medical imaging. Unlike a standard ROC curve, FROC allows for an unlimited number of marks per image, plotting sensitivity against the average number of false positives per scan. The confidence score directly determines where a model sits on this curve. A clinically viable CADe system might achieve 90% sensitivity at 0.5 false positives per image. FROC analysis is mandated in many FDA clearance studies for computer-aided detection devices.

The process of aligning a model's output confidence score with the true empirical likelihood of correctness. A perfectly calibrated model that outputs a confidence of 0.8 should have exactly 80% of those predictions be correct. Modern neural networks are notoriously miscalibrated, often producing overconfident scores. Techniques like temperature scaling—a single-parameter variant of Platt scaling applied to the logits—can dramatically improve calibration without affecting accuracy. In diagnostic AI, calibration is critical for risk stratification and clinical decision support.

An ensembling method that merges bounding boxes from multiple detection models by averaging their coordinates and confidence scores, rather than simply selecting one. Unlike NMS or soft-NMS which discard boxes, WBF uses all predictions from all models to construct a fused box. Each box's coordinates are weighted by its confidence score, meaning a high-confidence prediction from one model can pull the final box toward its location. This technique consistently outperforms standard NMS-based ensembling in medical imaging challenges where model diversity improves localization accuracy.