Small Object Detection

A foundational architecture that constructs a multi-scale, pyramidal hierarchy of feature maps with strong semantics at all levels. FPN fuses high-resolution, semantically weak features with low-resolution, semantically strong features via a top-down pathway and lateral connections.

• Mechanism: Builds feature maps at scales P2 through P7, where P2 has the highest resolution for detecting tiny objects.
• Clinical relevance: Essential for detecting micro-calcifications in mammography that may span only 3-5 pixels.
• Common pairing: Used as the backbone neck in Faster R-CNN and Mask R-CNN for radiological tasks.

Preserving the native resolution of medical scans is critical because down-sampling can entirely eliminate small lesions. Patch-based analysis tiles a gigapixel whole-slide image or large CT slice into overlapping smaller regions.

• Tiling strategy: Overlapping patches (e.g., 256x256 pixels with 50% overlap) ensure objects at patch boundaries are not missed.
• Resolution preservation: Mammography models often operate at 50-100 microns per pixel to retain micro-calcification morphology.
• Trade-off: Increases computational load linearly with the number of patches, requiring efficient batching.

Standard cross-entropy loss is overwhelmed by the vast number of easy negative examples in medical images. Focal Loss adds a modulating factor (1 - p_t)^γ to down-weight the loss contribution from well-classified examples.

• Formula: FL(p_t) = -α_t(1 - p_t)^γ log(p_t), where γ=2 is a common setting.
• Effect: Forces the model to focus training on hard, misclassified examples—typically the small, ambiguous lesions.
• Origin: Introduced in the RetinaNet paper and now standard in medical detection pipelines with severe foreground-background skew.

Dilated convolutions expand the receptive field of a kernel without increasing parameters or losing spatial resolution. By inserting holes (atrous) between kernel elements, the network captures broader contextual information while maintaining fine-grained localization.

• Benefit: A 3x3 kernel with dilation rate 2 has the receptive field of a 5x5 kernel but the parameters of a 3x3.
• Use case: Detecting early-stage lung nodules where surrounding tissue context is essential for distinguishing benign from malignant.
• Architecture: Often used in DeepLab-style segmentation heads adapted for small object detection.

A domain-specific augmentation where annotated small objects are copied from one image and pasted onto another training image at random locations. This artificially increases the density of rare positive examples.

• Implementation: Lesions are segmented, extracted, and blended onto new backgrounds using Poisson blending or simple alpha compositing.
• Impact: Dramatically improves recall for sub-5mm nodules by exposing the model to more varied contexts and backgrounds.
• Validation: Must ensure pasted objects do not violate anatomical constraints (e.g., a bone lesion on soft tissue).

Unlike Vision Transformers (ViT) that operate on coarse patches, Swin Transformers build hierarchical feature maps by merging image patches in deeper layers. Shifted window attention limits self-attention computation to local windows while enabling cross-window connections.

• Advantage: Produces multi-scale feature maps natively, similar to a ConvNet pyramid, making it ideal for small object detection.
• Clinical application: State-of-the-art performance on micro-metastasis detection in lymph node pathology.
• Efficiency: Linear computational complexity relative to image size, unlike the quadratic complexity of global self-attention.

A multi-scale feature extractor that builds a pyramidal hierarchy of feature maps, enabling detectors to recognize objects across vastly different sizes. FPN creates a top-down pathway with lateral connections, enriching shallow, high-resolution layers with semantic information from deeper layers.

• Critical for detecting objects smaller than 32×32 pixels
• Preserves spatial resolution in early layers for precise localization
• Used as the backbone in Faster R-CNN and RetinaNet for medical imaging
• Enables a single network to detect both micro-calcifications and large masses

A specialized loss function that addresses extreme class imbalance in dense object detection. Focal Loss dynamically down-weights the contribution of well-classified examples, forcing the model to focus training on hard, misclassified instances.

• Adds a modulating factor (1 - p_t)^γ to standard cross-entropy
• Typical γ value of 2.0 reduces easy-negative influence by orders of magnitude
• Essential for medical imaging where background pixels vastly outnumber small lesions
• Introduced in the RetinaNet architecture to match two-stage detector accuracy with single-stage speed

A quantization-free region pooling operation that preserves exact spatial locations when extracting features from Region of Interest proposals. Unlike RoI Pooling, RoI Align uses bilinear interpolation at sampled points to avoid misalignments caused by coordinate rounding.

• Critical for small objects where even 1-pixel misalignment destroys localization
• Enables accurate mask prediction in Mask R-CNN
• Preserves sub-pixel precision for micro-calcification boundaries
• Standard component in modern two-stage medical detection pipelines

The standard evaluation metric for object detection that calculates the mean of Average Precision scores across all classes at multiple IoU thresholds. For small object detection, mAP@[0.5:0.95] with small-area designation (objects < 32² pixels) is the definitive benchmark.

• AP_small specifically measures performance on objects under 32×32 pixels
• COCO-style evaluation uses 10 IoU thresholds from 0.50 to 0.95
• Medical imaging often uses [email protected] for clinical relevance
• Balances precision-recall trade-off across all operating points

An evaluation metric designed specifically for medical detection tasks that plots sensitivity against the average number of false positives per image. Unlike mAP, FROC allows an unlimited number of marks per scan, making it clinically relevant for CADe systems.

• X-axis: false positives per image (e.g., 0.125, 0.25, 0.5, 1, 2, 4, 8)
• Y-axis: lesion-level sensitivity at each operating point
• Standard metric for mammography CADe and lung nodule detection challenges
• Reflects radiologist workflow tolerance for false alarms

A modern detection paradigm that directly predicts object centers or keypoints without relying on predefined anchor boxes. Methods like CenterNet and FCOS eliminate anchor-related hyperparameters, simplifying small object detection where anchor scale selection is notoriously difficult.

• Predicts a heatmap of object centers with regression to bounding box dimensions
• Avoids the anchor-box mismatch problem for objects of unusual aspect ratios
• Reduces computational overhead from dense anchor matching
• Particularly effective for detecting small, circular lesions like lung nodules