BI-RADS Consistency Regularization (BCR)

• BCR is a domain-informed regularization strategy that embeds BI-RADS priors into deep models for improved breast lesion segmentation and diagnosis.
• It enforces morphological consistency and feature-space alignment, mitigating destructive task interference in multi-task and longitudinal imaging.
• Empirical validations demonstrate significant performance gains, including higher Dice scores and better generalization across diverse breast imaging datasets.

BI-RADS Consistency Regularization (BCR) encompasses a family of domain-informed regularization strategies that inject Breast Imaging-Reporting and Data System (BI-RADS) prior knowledge into deep learning models for breast lesion segmentation and classification. BCR ameliorates destructive task interference in multi-task learning, improves domain generalization, and aligns latent feature representations with radiologic assessment by linking model predictions or embeddings to clinically meaningful BI-RADS-derived morphologic or diagnostic information. Notably, BCR has been instantiated as a morphological consistency term in externally validated breast ultrasound segmentation (Zhang et al., 20 Nov 2025), and as a feature-space regularizer for longitudinal breast MRI sequence modeling (Kamran et al., 1 Aug 2025).

1. Conceptual Foundation and Motivation

BCR arises from the recognition that multi-task and longitudinal learning tasks in breast imaging frequently underperform compared to single-task or single-timepoint alternatives, due to destructive interference between task losses or neglect of clinical context. In breast ultrasound segmentation, simultaneously predicting malignancy and spatial masks can degrade segmentation accuracy, particularly on external datasets. Incorporating BI-RADS-inspired morphologic consistency constrains the model to respect known structure—such as irregular boundary, area, and texture cues—common in radiologic diagnosis (Zhang et al., 20 Nov 2025).

In longitudinal breast MRI, BI-RADS scores over time delineate stable versus evolving findings; therefore, BCR leverages this to regularize latent representations: embedding similarity is enforced when BI-RADS remains stable, and divergence is permitted or encouraged under diagnostic change (Kamran et al., 1 Aug 2025). This operationalizes the clinical workflow where radiologists interpret imaging in the context of prior assessments.

2. Mathematical Formulations and Loss Structures

BCR is instantiated as an auxiliary loss term, tightly coupled to clinical priors. Two principal variants have been introduced:

A. Morphological Consistency (Ultrasound, (Zhang et al., 20 Nov 2025)):

• Given input $x$, the network outputs segmentation logits $s(x)$ and a malignancy logit $m(x)$.
• Four differentiable morphological features (Area $A$, Roughness $R$, Compactness $1-C$, Texture $T$) are computed from the segmentation mask.
• A composite prior $\varphi = w^{\top}f$, with $w = \operatorname{softmax}(u)$ initialized to represent BI-RADS heuristic weights and trained end-to-end.
• The consistency loss is $$L_{\text{cons}} = \operatorname{MSE}(\hat{p}, \varphi)$$, where $s(x)$0 is the malignancy probability.
• The full objective:

$s(x)$1

with task-specific weights, auxiliary penalties, and explicit regularization of BI-RADS parameters.

B. Feature-Space Consistency (MRI Longitudinal, (Kamran et al., 1 Aug 2025)):

• For paired current/prior scans $s(x)$2 and associated BI-RADS scores:
• At each U-Net skip-connection layer $s(x)$3,

$s(x)$4

• The total regularizer is $s(x)$5 with $s(x)$6 emphasizing deeper layers.
• The BCR term blends with Dice and cross-entropy segmentation losses, tuning their balance for task-specific networks.

3. Algorithmic Instantiation and Integration

BCR for Breast Ultrasound Multi-task Segmentation (Zhang et al., 20 Nov 2025)

BCR is injected as a bridging penalty between the segmentation output and the malignancy classifier:

• Network employs an EfficientNet-B1 encoder + U-Net style decoder with dual-stream input (3-channel for encoder, grayscale for texture).
• Morphological BI-RADS-derived features are computed directly from the segmentation output at each forward pass.
• After forming the malignancy prior $s(x)$7, the MSE consistency loss is back-propagated through both segmentation and classification heads, as well as the differentiable feature-computation graph.

BCR for Longitudinal MRI Representation Alignment (Kamran et al., 1 Aug 2025)

BCR is employed in a dual-encoder 3D U-Net processing current and previous DCE-MRI timepoints:

• At each skip-connection layer, Euclidean distance between feature maps is penalized inversely proportional to the change in BI-RADS label.
• The loss only applies during training (requiring BI-RADS annotations for prior and current images); at inference, no loss calculation or additional annotation is necessary.
• Layer-specific weights increase with semantic depth, focusing regularization on high-level features.

4. Empirical Evaluation and Impact

Externally Validated Generalization (Zhang et al., 20 Nov 2025)

Comprehensive external validation was conducted, training on the BrEaST (Poland) dataset and evaluating without fine-tuning on BUSI (Egypt), UDIAT (Spain), and BUS-UCLM (Spain):

Dataset	seg_only	seg_malig	Proposed (BCR)	% Gain over seg_malig	p-value
BUSI	0.65	0.56	0.66	+18%	<0.001
UDIAT	0.78	0.59	0.81	+37%	<0.001
BUS-UCLM	0.64	0.49	0.69	+41%	<0.001

BCR recovers and surpasses single-task Dice under domain shift, yielding state-of-the-art performance on UDIAT.

LesiOnTime: Small Lesion Segmentation in DCE-MRI (Kamran et al., 1 Aug 2025)

On a curated high-risk patient cohort:

• LesiOnTime without BCR: Dice = 0.32.
• LesiOnTime with BCR: Dice = 0.35 (~9% gain).
• Hausdorff-95 distance improved (112.3 mm to 106.5 mm).
• Embedding ablation shows that BCR clusters feature vectors for BI-RADS-stable timepoints and allows separation when the assessment evolves.

5. Ablation Studies and Feature Importance

Ablation analyses in (Zhang et al., 20 Nov 2025) demonstrate that removing any single morphological descriptor (area, roughness, compactness, texture) from $s(x)$8 leads to a 3–5% drop in external Dice, indicating all features contribute additively to robustness. Qualitative review confirms that BCR-trained models delineate spiculated or low-contrast tumor boundaries with improved precision and reduced noise.

Similarly, ablating BCR from the LesiOnTime framework results in increased embedding distances for timepoint pairs with identical BI-RADS, disrupting the desired latent alignment. This supports the claim that BCR not only enhances segmentation but also imposes clinically meaningful structure on learned representations (Kamran et al., 1 Aug 2025).

6. Limitations, Extensions, and Generalization

Current BCR studies are limited to center-specific datasets and pre-defined task or layer weighting schemes, with weights empirically tuned or heuristically scheduled. Possible extensions include learning the task weights or denominator scaling, employing alternative distance metrics (e.g., cosine or Mahalanobis), or extending regularization to additional clinical scores or biomarkers.

A plausible implication is that BCR-style methodology may improve generalization and interpretability for deep networks in other radiological tasks where domain priors and structured reporting are available.

7. Summary and Perspective

BI-RADS Consistency Regularization operationalizes radiological expertise by fusing structured domain priors—either as morphological features or as temporal diagnostic trajectories—directly into deep learning objectives. In externally validated breast ultrasound segmentation, BCR achieves statistically significant improvements in Dice score under considerable domain shift, mitigates destructive task interference, and maintains competitive malignancy classification. In longitudinal MRI, BCR yields improved segmentation and alignment of latent representations with radiologic interpretation, even for subtle, small lesions (Zhang et al., 20 Nov 2025, Kamran et al., 1 Aug 2025).

BCR represents a paradigm in which domain-driven, clinically grounded regularization is directly encoded within algorithmic training, enabling improved robustness and clinical validity for diagnostic AI in breast imaging.