Computer-Aided Diagnosis (CADx)

CADx systems output a continuous probability score (0.0 to 1.0) representing the likelihood of malignancy, rather than a binary detection mark. This score is derived from a softmax activation layer in the classification head. The raw logit is often calibrated using Platt scaling or isotonic regression to ensure the predicted confidence aligns with empirical observed frequencies. A score of 0.95 indicates the model is 95% confident the lesion is malignant, enabling risk-stratified clinical decision-making.

A core function of CADx is the automated mapping of imaging features to the Breast Imaging Reporting and Data System (BI-RADS) lexicon. The model learns to assign a standardized category (0–6) based on morphological characteristics:

• BI-RADS 3: Probably benign (< 2% malignancy risk)
• BI-RADS 4: Suspicious abnormality (subdivided into 4A, 4B, 4C)
• BI-RADS 5: Highly suggestive of malignancy (≥ 95% risk) This structured output ensures seamless integration with clinical reporting workflows and audit trails.

CADx engines quantify specific radiological features that correlate with histopathological outcomes. Key extracted features include:

• Spiculation: Radiating lines from a mass margin, a highly specific indicator of invasive ductal carcinoma.
• Margin sharpness: Ill-defined vs. circumscribed boundaries.
• Internal echo pattern: Heterogeneous vs. homogeneous texture.
• Calcification morphology: Pleomorphic vs. punctate vs. amorphous. These features are often processed by parallel attention heads in a Vision Transformer architecture before fusion.

Advanced CADx systems integrate data from multiple imaging modalities and clinical sources to improve diagnostic accuracy. A late-fusion architecture might concatenate embeddings from:

• FFDM or DBT images for structural analysis.
• Contrast-Enhanced Mammography (CEM) for kinetic curve assessment.
• Patient metadata (age, family history, genetic risk). The fused representation is passed through a gated attention mechanism that learns to weight each modality based on its diagnostic contribution for the specific case.

Beyond a single malignancy score, some CADx systems generate a ranked differential diagnosis list. The model outputs a probability distribution over multiple pathology classes:

• Invasive Ductal Carcinoma (IDC)
• Ductal Carcinoma In Situ (DCIS)
• Fibroadenoma
• Complex Sclerosing Lesion (Radial Scar)
• Intramammary Lymph Node This is achieved using a multi-class cross-entropy loss during training, providing the radiologist with a structured set of diagnostic hypotheses rather than a single opaque score.

To support clinical trust and regulatory compliance, CADx systems incorporate post-hoc explainability techniques. Methods like Grad-CAM or Integrated Gradients generate a heatmap overlay highlighting the specific pixels and morphological structures that most influenced the malignancy score. This allows the radiologist to visually verify that the model's reasoning aligns with clinically relevant findings (e.g., focusing on a spiculated margin) rather than spurious correlations or image artifacts.

The essential precursor to CADx. CADe algorithms scan the entire mammogram to localize Regions of Interest (ROIs) that contain potential abnormalities, such as masses or microcalcifications. While CADe answers 'where is the lesion?', CADx answers 'what is the lesion?' Modern deep learning pipelines often fuse these two stages into a single end-to-end architecture, but regulatory frameworks frequently evaluate them as distinct modules.

The Breast Imaging Reporting and Data System (BI-RADS) is the standardized lexicon that CADx models are trained to predict. A CADx system maps extracted imaging features to a numerical category:

• BI-RADS 1 & 2: Negative or benign findings.
• BI-RADS 3: Probably benign, suggesting short-interval follow-up.
• BI-RADS 4 & 5: Suspicious abnormality or highly suggestive of malignancy, warranting biopsy. Accurate BI-RADS assignment is the primary regulatory endpoint for CADx software.

Pixel-level lesion segmentation is a critical input for CADx feature extraction. By precisely delineating the boundary of a mass from surrounding parenchyma, the system can compute quantitative morphological descriptors such as margin sharpness, spiculation, and shape irregularity. Architectures like U-Net or Mask R-CNN generate the binary masks that feed downstream characterization classifiers.

CADx engines often rely on radiomics, the high-throughput extraction of quantitative features from the segmented ROI. These include:

• First-order statistics: Histogram analysis of pixel intensities.
• Shape features: Compactness, sphericity, and surface-to-volume ratio.
• Texture matrices: Gray-Level Co-occurrence Matrix (GLCM) and Run-Length Matrix (GLRLM) to quantify tissue heterogeneity. These handcrafted features are frequently combined with deep features in hybrid models.

A critical post-processing step ensuring that a CADx model's output probability genuinely reflects empirical risk. A well-calibrated model will have a malignancy probability of 0.8 that corresponds to an 80% observed malignancy rate. Techniques like Platt scaling or isotonic regression are applied to raw logits to correct overconfidence, which is essential for clinical trust and integration into the BI-RADS framework.

Advanced CADx systems achieve superior accuracy through multi-modal fusion, integrating mammographic findings with non-imaging data. Inputs can include:

• Patient demographics and age.
• Genetic risk factors (e.g., BRCA status).
• Prior clinical history.
• Contrast-enhanced kinetics. Fusion architectures use cross-attention mechanisms to weigh contradictory evidence, mimicking a radiologist's holistic diagnostic synthesis.