Class Imbalance

A loss function designed to address extreme class imbalance by dynamically scaling the cross-entropy loss. It down-weights the loss assigned to well-classified, easy examples, forcing the model to focus on hard, misclassified examples.

• Mechanism: Adds a modulating factor (1 - p_t)^γ to the standard cross-entropy loss.
• Effect: Prevents the vast number of easy negative examples from overwhelming the detector during training.
• Hyperparameter: The focusing parameter γ adjusts the rate at which easy examples are down-weighted; γ=2 is standard.

A training strategy that explicitly identifies and re-trains on false positive detections. Instead of sampling randomly, the model iteratively finds background patches that it incorrectly classified as objects with high confidence.

• Online Hard Example Mining (OHEM): Selects the highest-loss RoIs in a forward pass and uses only those for backpropagation.
• Benefit: Directly reduces the false alarm rate by forcing the model to learn the distinction between difficult background textures and true pathologies.

A pre-processing approach that modifies the training dataset distribution before learning begins.

• Oversampling: Duplicates rare positive examples (e.g., lesions) to increase their presence in a batch. Risk of overfitting.
• Undersampling: Randomly discards a large portion of the majority class (background patches) to balance the ratio.
• Class Weights: Assigns an inverse-frequency weight to each class in the loss function, penalizing mistakes on the rare class more heavily.

Generates new, artificial positive examples rather than simply duplicating existing ones. In medical imaging, this often involves data augmentation specific to the pathology.

• Techniques: Random rotations, elastic deformations, and intensity shifts applied exclusively to the minority class bounding boxes.
• Advanced: Generative Adversarial Networks (GANs) can synthesize entirely new, realistic lesion appearances to augment the training set without compromising patient privacy.

Architectures like Faster R-CNN inherently mitigate imbalance through their design. The first stage (Region Proposal Network) filters the vast majority of background regions, passing only a sparse set of candidate locations to the second stage.

• Ratio Control: The second stage classifier is trained with a fixed foreground-to-background ratio (e.g., 1:3), ensuring a balanced mini-batch.
• Contrast: Single-stage detectors (YOLO, SSD) must process the entire dense image grid, making them more susceptible to imbalance without Focal Loss.

Standard accuracy is a misleading metric for imbalanced medical data. A model predicting 'no tumor' on every scan achieves 99.9% accuracy but zero clinical utility.

• Precision-Recall Curve: Focuses exclusively on the positive class performance.
• FROC (Free-Response ROC): The standard for radiological CADe, plotting sensitivity against the average number of false positives per image.
• mAP: Evaluates precision and recall at various IoU thresholds, providing a holistic view of detection quality.

A training strategy that explicitly identifies and re-trains on false positive detections (hard negatives) to reduce the model's false alarm rate. Instead of randomly sampling background patches, the model iteratively mines the specific background regions that are most frequently confused with foreground objects.

• Online Hard Example Mining (OHEM): Automatically selects the highest-loss RoIs in each batch
• Benefit: Prevents the gradient from being dominated by trivial background classification
• Application: Critical for reducing false positives in lung nodule detection

Techniques that artificially modify the class distribution of the training dataset to prevent model bias toward the majority class. These operate at the data-loading level before the loss function.

• Oversampling: Duplicating rare positive examples (e.g., lesions) to increase their representation
• Undersampling: Randomly discarding background patches to balance the batch composition
• Risk: Oversampling can lead to overfitting; undersampling can discard useful contextual information

An evaluation metric specifically designed for detection tasks with extreme class imbalance, where the number of false positives per image is more clinically relevant than pixel-level accuracy. It plots sensitivity against the average number of false positives per scan.

• Axes: Y-axis is recall rate; X-axis is false positives per image (FPPI)
• Advantage: Allows an unlimited number of marks per scan, unlike standard ROC
• Clinical Relevance: Measures how many false alarms a radiologist must dismiss per case

A programmatic approach to class imbalance where each training sample is assigned a probability of being selected based on its class frequency. Rare classes receive higher sampling weights, ensuring each mini-batch contains a balanced mix of foreground and background.

• Implementation: PyTorch WeightedRandomSampler or TensorFlow class_weight
• Granularity: Can be applied at the image level or the instance level
• Best Practice: Combine with heavy data augmentation on the oversampled rare classes to prevent memorization

An algorithmic oversampling technique that creates synthetic positive examples rather than simply duplicating existing ones. It generates new samples by interpolating between a rare example and its nearest neighbors in feature space.

• Process: Selects a sample, finds k-nearest neighbors, and interpolates a new point
• Limitation: Operates in feature space, not pixel space; less common for raw images
• Alternative: Use GANs or diffusion models for synthetic lesion generation in pixel space