Intersection over Union (IoU)

IoU is inherently scale-invariant, meaning the metric value remains identical regardless of the absolute size of the objects being compared. Whether evaluating a 20-pixel micro-calcification or a 500-pixel lung mass, the ratio of overlap to union is unaffected by the object's pixel area. This property is critical in medical imaging, where pathologies span orders of magnitude in size.

• Why it matters: A model's performance on small lesions and large tumors can be compared directly without normalization.
• Contrast: Pixel-wise accuracy metrics like L2 distance are scale-dependent and penalize large objects disproportionately.

IoU is strictly bounded between 0.0 and 1.0, where:

• 0.0 indicates zero overlap between the predicted and ground truth boxes.
• 1.0 represents perfect, pixel-level agreement.

This bounded interval makes IoU immediately interpretable as a percentage of spatial agreement. In object detection benchmarks, an IoU threshold of 0.5 is commonly used for a true positive match, while 0.75 defines a strict, high-quality localization standard.

IoU simultaneously penalizes two distinct failure modes in a single metric:

• False Positives: Pixels predicted as the object that are not in the ground truth annotation increase the union area without increasing the intersection, driving IoU toward 0.
• False Negatives: Pixels in the ground truth that the model missed also reduce the intersection while the union remains large.

This dual penalization makes IoU more informative than simple pixel accuracy, which can be misleadingly high when a model predicts the entire image as background in a sparse detection task.

For a single discrete prediction, IoU is a non-differentiable function because it depends on hard thresholding of predicted bounding box coordinates. This presents a challenge for direct optimization via gradient descent.

• Workaround: Modern detection architectures use smooth L1 loss or GIoU (Generalized IoU) as differentiable surrogate loss functions during training.
• Evaluation only: Standard IoU is reserved strictly as an evaluation metric, while its differentiable approximations guide the learning process.

In object detection evaluation, IoU serves as a gating function for true positive classification. A predicted box is considered a correct detection only if its IoU with a ground truth box exceeds a predefined threshold.

• [email protected]: Mean Average Precision at IoU threshold 0.5, the most lenient standard.
• [email protected]:0.95: The COCO challenge standard, averaging mAP across IoU thresholds from 0.5 to 0.95 in 0.05 increments, rewarding models with tight, precise localization.
• Medical imaging often uses stricter thresholds (e.g., 0.7) for clinical acceptance.

IoU is a symmetric metric, meaning:

IoU(A, B) = IoU(B, A)

Swapping the predicted box and the ground truth box yields the identical score. This symmetry ensures that the metric treats both inputs equally and does not bias evaluation toward either the prediction or the annotation. This property distinguishes IoU from asymmetric metrics like precision and recall, which treat the prediction as the primary subject and the ground truth as the reference.

The de facto standard for summarizing object detector performance across all classes and IoU thresholds.

• Calculates Average Precision (AP) per class by integrating the Precision-Recall curve
• [email protected] evaluates at a single, lenient IoU threshold of 0.5
• mAP@[0.5:0.95] averages AP across 10 thresholds from 0.5 to 0.95 in 0.05 increments, rewarding precise localization
• Essential for comparing models on benchmarks like COCO and medical imaging challenges

A loss function engineered to combat the extreme foreground-background class imbalance inherent in dense object detection.

• Down-weights the loss contribution from well-classified examples (easy negatives)
• Introduces a focusing parameter γ that reduces the relative loss for examples with high confidence
• Forces the model to concentrate training on hard, misclassified examples
• Critical for detecting rare pathologies where background pixels vastly outnumber lesion pixels

A post-processing algorithm that eliminates redundant, overlapping detections for the same object.

• Sorts all predicted boxes by confidence score
• Selects the highest-confidence box and suppresses all others with IoU > threshold
• Soft-NMS decays scores instead of hard-suppressing, improving recall for occluded or clustered abnormalities
• Directly uses IoU as the overlap criterion for suppression decisions

The mechanism by which a detector refines coarse region proposals into precise localization coordinates.

• Predicts offsets (Δx, Δy, Δw, Δh) relative to anchor boxes or initial proposals
• Smooth L1 Loss is commonly used as it is less sensitive to outliers than L2 loss
• IoU Loss directly optimizes the IoU metric itself, improving alignment between the optimization objective and evaluation criterion
• GIoU/DIoU/CIoU Loss extend IoU loss to handle non-overlapping boxes and incorporate aspect ratio

An evaluation metric specifically designed for medical detection tasks where multiple findings per image are expected.

• Plots sensitivity against the average number of false positives per scan
• Unlike ROC, allows an unlimited number of marks per image
• A FROC curve reaching high sensitivity at low false-positive rates indicates clinical viability
• Commonly used in mammography CADe and lung nodule detection challenges

A multi-stage detection architecture that trains a sequence of detectors with progressively higher IoU thresholds.

• Stage 1: IoU threshold 0.5, Stage 2: 0.6, Stage 3: 0.7
• Each stage refines the bounding boxes from the previous stage
• Addresses the overfitting problem caused by a single IoU threshold during training
• Produces high-quality, tightly localized boxes critical for measuring lesion dimensions accurately