CADe (Computer-Aided Detection)

The core function of a CADe system is the automated localization of potential abnormalities. Unlike manual screening, the system algorithmically scans the entire image volume and places bounding boxes or visual prompts over areas with statistically anomalous patterns. This process is designed to combat satisfaction of search, a perceptual error where a radiologist stops looking after finding the first abnormality, by ensuring all suspicious loci are highlighted for review.

CADe algorithms are explicitly tuned for maximum sensitivity, often at the expense of generating false positives. The clinical goal is to achieve a near-100% detection rate for actionable pathologies, ensuring no true lesion is missed. This design philosophy accepts a higher false-positive-per-image rate because the cost of a missed cancer far outweighs the radiologist's cognitive load of dismissing a false mark. The system acts as a safety net, not a definitive diagnostician.

To detect abnormalities of vastly different sizes—from a 2mm micro-calcification to a large mass—CADe systems employ multi-scale architectures like Feature Pyramid Networks (FPN). These networks build a pyramidal hierarchy of feature maps, allowing the model to leverage high-resolution, semantically weak features for small objects and low-resolution, semantically strong features for large objects. This ensures robust detection across the entire size spectrum of potential findings.

Raw neural network outputs often produce multiple overlapping detections for a single lesion. CADe systems apply Non-Maximum Suppression (NMS) or Weighted Boxes Fusion (WBF) to consolidate redundant bounding boxes into a single, precise mark. This post-processing step is critical for clinical usability; without it, a single mass could be obscured by a cluster of marks, increasing cognitive burden rather than reducing it.

For seamless clinical workflow integration, CADe outputs must conform to the DICOM Structured Reporting (DICOM SR) standard. This encodes detection results—including bounding box coordinates, measurements, and confidence scores—into a structured format that can be ingested by a Picture Archiving and Communication System (PACS). This allows the CADe marks to be overlaid directly on the original scan in the radiologist's standard viewing environment.

A significant engineering challenge is the domain shift caused by different scanner manufacturers and acquisition protocols. A CADe system trained on Siemens scans may fail on GE scans. Robust systems employ domain adaptation techniques, such as adversarial training or Hounsfield Unit (HU) normalization, to learn scanner-invariant features. This ensures consistent sensitivity regardless of the imaging hardware used to acquire the scan.