BI-RADS-Inspired Morphological Features

• BI-RADS-inspired morphological features are quantitative descriptors that emulate radiologist BI-RADS lexicon for lesion characterization in breast imaging.
• They integrate mathematical formulations and radiomic metrics with classical and deep learning pipelines to significantly boost diagnostic accuracy and reproducibility.
• These features enable explainable AI by mapping computed descriptors to clinical semantics, thereby improving interpretability and external validation in CAD systems.

BI-RADS-inspired morphological features are quantitative or learnable image descriptors explicitly designed to emulate the lesion characterizations found in the American College of Radiology Breast Imaging Reporting and Data System (BI-RADS) lexicon. These features bridge the gap between subjective radiologist annotation and algorithmic pattern recognition, underpinning explainable, reproducible, and generalizable computer-aided diagnosis (CAD) across a range of breast imaging modalities. The following sections survey the core classes of BI-RADS-inspired morphological features, their mathematical and algorithmic instantiations, integration within traditional and deep learning pipelines, empirical contributions to diagnostic accuracy, and role in interpretability and external validation.

1. Canonical BI-RADS Morphological Descriptors

BI-RADS specifies several morphological categories foundational to breast lesion risk assessment:

• Shape: round, oval, or irregular.
• Margin: circumscribed, indistinct, microlobulated, angular, or spiculated.
• Orientation: parallel (“wider-than-tall”) vs. not-parallel (“taller-than-wide”).
• Echo Pattern: anechoic, hypoechoic, isoechoic, hyperechoic, complex cystic-solid, heterogeneous.
• Posterior Features: none, enhancement, shadowing, combined pattern.

Architectures such as BI-RADS-Net (Zhang et al., 2021) and MT-BI-RADS (Karimzadeh et al., 2023) explicitly predict these descriptors via multi-head classification schemes—including both primary categories and margin subtypes—enabling parallel extraction of the radiologist taxonomy and core diagnostic output (benign/malignant). These descriptors are mapped either to categorical labels (via softmax/sigmoid heads) or, in advanced CAD, further quantified by geometric morphometrics and radiomics.

2. Mathematical Formulation of Morphological Features

Handcrafted BI-RADS-inspired features numerically refine the morphological lexicon, offering reproducibility and objective thresholds unattainable by raw categorical coding. Multiple studies (Byra et al., 2017, Gorji et al., 21 Jul 2025, Ardakani et al., 31 Aug 2025, Boumaraf et al., 2020) provide formulas and paradigms. Salient examples include:

Descriptor	Formula	BI-RADS Mapping
Circularity	$C = \frac{4\pi A}{P^2}$	round/oval vs. irregular
Ellipticity	$E = \frac{A}{A_{ellipse}}$	oval vs. irregular
Convexity	$Conv = P_{convex-hull}/P$	margin regularity
Extent	$Extent = A/A_{bbox}$	compactness
Overlap Ratio	$OR = A_h/A$	boundary lobulation
NRL Entropy	$H = -\sum p_k\log p_k$ (radial length histogram)	margin undulation/spiculation
Aspect Ratio	$AR = L_{major}/L_{minor}$	shape/orientation
Depth-to-Width (DWR)	$DWR = height/width$	orientation
Spiculation	$(P-P_h)/P_h$ or high-freq NRL count	margin spiculation

Higher-dimensional radiomics further encode 3D sphericity, flatness, elongation, surface-to-volume ratio, and axis lengths, directly quantifying roundness, compactness, and complexity in ways that emulate but surpass visual descriptors (Gorji et al., 21 Jul 2025, Salmanpour et al., 14 Dec 2024).

3. Integration into Machine Learning and Deep Learning Pipelines

Morphological features are utilized via two main strategies:

• Handcrafted Feature Extraction and Classical ML: Morphometric features are explicitly computed on segmented ROIs and fed into classifiers (e.g., logistic regression, SVMs, shallow neural nets) (Byra et al., 2017, Boumaraf et al., 2020, Ardakani et al., 31 Aug 2025, Gorji et al., 21 Jul 2025). Dimensionality is controlled via feature selection (e.g., LASSO, stepwise selection, genetic algorithms), maximizing AUC and interpretability.
• Feature-Driven Deep Learning: Recent networks integrate BI-RADS features in both the input preprocessing stage and as auxiliary tasks or differentiable constraints. The BIRADS-SDL architecture (Zhang et al., 2019) introduces a “BIRADS-oriented feature map” (BFM) preprocessing using boundary distance transform with Gaussian weighting:

$$BFM(p)=I(p)\cdot\exp\left(-\frac{\text{Dist}(p)^2}{\sigma^2}\right)$$

This enhances boundary-proximal structures and guides the encoder to focus on clinically relevant topologies. Further, differentiable proxies for area, roughness (via Sobel gradient), compactness, and echotexture are directly imposed as soft consisitency regularizers tying segmentation and classification outputs (Zhang et al., 20 Nov 2025). Multi-task frameworks learn descriptor heads alongside lesion classification, further regularized by agreement or consistency losses (Zhang et al., 2021, Karimzadeh et al., 2023).

4. Diagnostic Impact and Empirical Results

Integration of BI-RADS-inspired features demonstrably enhances performance across modalities and tasks:

• CAD Performance: Combined BI-RADS + morphometric classifiers robustly outperform either set alone. For example, in breast ultrasound, the fused classifier of classic BI-RADS and six morphological features delivered an AUC of 0.986, sensitivity of 96.8%, and specificity of 94.7%, substantially above BI-RADS-only (AUC = 0.944) or morphometry alone (AUC = 0.901) (Byra et al., 2017, Ardakani et al., 31 Aug 2025).
• Deep Learning Gains: Embedding feature maps or descriptor heads into deep networks such as BIRADS-SDL, BI-RADS-Net, and MT-BI-RADS improve classification accuracy (up to ~92% in US, 91.3% in BUS) and segmentation Dice (up to 0.81 in external validation) (Zhang et al., 2019, Zhang et al., 2021, Karimzadeh et al., 2023, Zhang et al., 20 Nov 2025). Incorporation of handcrafted text-based descriptors into transformer-style architectures raises AUC from 0.711→0.872 (∆+0.161) in mammogram diagnosis compared to image-only baselines (Ben-Artzi et al., 16 Nov 2024).
• External Validation and Generalizability: Consistency-regularized multi-task networks using differentiable morphological features eliminate destructive task interference and result in substantial generalization improvements (e.g., +37% Dice on UDIAT dataset) over naive baselines (Zhang et al., 20 Nov 2025).

5. Interpretable AI and Clinical Transparency

BI-RADS-inspired features serve as anchors for explainable predictions in clinical settings:

• Descriptor-Level Explanations: Multi-task architectures furnish explicit predictions for each BI-RADS category, closely mirroring the clinical reporting language and mapping learned features to human-interpretable semantics (Zhang et al., 2021, Karimzadeh et al., 2023).
• Global and Local Attribution: Post-hoc explanations via SHAP quantify the individual descriptor’s impact on a given malignancy decision (e.g., φ_spiculated = +0.20, φ_parallel = –0.30) (Karimzadeh et al., 2023). In radiomics pipelines, SHAP analysis links feature importance (e.g., sphericity, perimeter-area ratio, convexity) directly to their BI-RADS analogs, confirming clinical intuition (Gorji et al., 21 Jul 2025).
• Nomogram and LLM Integration: Morphometric nomograms integrating BI-RADS and quantitative features achieved higher accuracy in biopsy decision-making than both radiologists and LLMs, supporting operationalizable, interpretable CAD (Ardakani et al., 31 Aug 2025).

6. Extensions to Other Modalities and Organ Systems

The abstraction of BI-RADS-inspired morphological quantification is not limited to breast imaging. Dictionary frameworks such as BM1.0 for breast MRI and PM1.0 for prostate MRI formalize analogous shape metrics—major/minor/least axis lengths, surface-area-to-volume ratio, sphericity, compactness, flatness, elongation—with precise LaTeX definitions and statistical associations to malignancy risk (Gorji et al., 21 Jul 2025, Salmanpour et al., 14 Dec 2024). This generalization creates a standardized quantitative bridge between radiological “lexica” (e.g., PI-RADS) and radiomics-driven computer vision pipelines.

7. Limitations and Open Challenges

Despite substantial progress, several unresolved issues persist:

• Segmentation Dependence: Most morphological quantification relies on accurate lesion mask extraction, which can be susceptible to operator bias or algorithmic error (Zhang et al., 2019).
• Domain Transfer: Numeric thresholds and feature distributions are scanner- and population-dependent; generalizability requires robust normalization and external validation (Ardakani et al., 31 Aug 2025, Zhang et al., 20 Nov 2025).
• Integration Complexity: Multi-modal fusion and consistency constraints impose nontrivial training dynamics (risk of task interference, overfitting) which require carefully tuned loss-weighting and joint optimization (Zhang et al., 20 Nov 2025).
• Subjective-Objective Mapping: While many features map well to BI-RADS definitions, some (e.g., internal complexity) have less direct clinical correlates, requiring careful interpretation and user education (Gorji et al., 21 Jul 2025).

A plausible implication is that ongoing development of standardized feature dictionaries, task-aligned regularization (e.g., via differentiable proxies), and post-hoc explanation pipelines will be central to both domain generalization and regulatory/clinical trust in AI-augmented breast imaging.

---

References:

• (Zhang et al., 2019, Byra et al., 2017, Zhang et al., 20 Nov 2025, Zhang et al., 2021, Karimzadeh et al., 2023, Gorji et al., 21 Jul 2025, Ben-Artzi et al., 16 Nov 2024, Salmanpour et al., 14 Dec 2024, Ardakani et al., 31 Aug 2025, Boumaraf et al., 2020)