Reading the Mask: Descriptor-Grounded BI-RADS Rationale Generation for Breast MRI Segmentation

• Segmentation masks provide lesion boundaries but not the image-feature explanations radiologists use to assess malignancy risk, limiting interpretability and trust.Source: §1 · p. 1
• Grad-CAM [19] highlights influential regions without clinical descriptors, while report-supervised PI-RADS reasoning depends on costly, difficult-to-obtain radiologist reports.Source: §1 · p. 1
• The framework derives BI-RADS descriptors from masks and dynamic MRI measurements, then generates rationales without additional radiologist annotation across 1,506 MAMA-MIA cases.It evaluates faithfulness, hallucination, omission, counterfactual accuracy, and descriptor changes between expert and automatic masks.Source: §1 · p. 1

• Gradient-based methods [19] highlight pixels without specifying a clinical rationale, while concept bottlenecks [9] may not match their named concepts [13].Source: §2 · p. 2
• Prototype classifiers [2] are limited to learned examples, and global image-caption alignment such as BiomedCLIP [21] cannot distinguish multiple findings in one organ.Source: §2 · p. 2
• Quantitative DCE-MRI features have been related to BI-RADS descriptors [6], and kinetic curves classified automatically; template-based chest X-ray reports have also been generated.Source: §2 · p. 2

• MAMA-MIA contains 1,506 studies assembled from the ISPY2, DUKE, ISPY1, and NACT cohorts.Source: §3 · p. 2
• The dataset includes expert masks and autogenerated masks visually graded by two experts as Good, Acceptable, Poor, or Missed.Source: §3 · p. 2
• Studies have three to six post-contrast phases and acquisition timestamps; a multifocality indicator is available for 1,346 cases.Source: §3 · p. 2

4.1 Step 1: Preparing the Lesion Representation

• For multifocal masks, the largest connected component supplies morphological measurements, while total component count represents multiplicity.The lesion region is resampled to isotropic 1 mm3 spacing before surface-based features are computed, addressing variation in voxel spacing.Source: §4.1.1 · p. 2, §4.1 · p. 2, §4.1.2 · p. 2

4.2.1 Morphological measurements.

• Four morphological measurements quantify lesion shape and boundary structure, including sphericity, elongation, surface-to-volume ratio, and compactness.Surface-to-volume ratio serves as a proxy for the BI-RADS margin descriptor, though it also depends on lesion size.Source: §4.2.1 · pp. 2–3

4.2.2 Contrast-enhancement measurements.

• Contrast-enhancement measurements use pre-contrast, first post-contrast, and final post-contrast images to characterize spatial and temporal uptake.SER distinguishes washout, plateau, and persistent or progressive enhancement; a too-small eroded core leaves internal enhancement unassessed.Source: §4.2 · p. 2, §4.2.2 · p. 3

4 PROPOSED APPROACH

• The pipeline maps eight MRI- and mask-derived measurements, lesion size, and component count to BI-RADS descriptors, then verbalizes the structured record as a rationale.The measurement set covers lesion morphology, internal enhancement, and temporal enhancement.Source: §4 · p. 2, §4.2.2 · p. 4

4.3.1 BI-RADS decision tree.

• The BI-RADS decision tree computes eight of 14 nodes from MRI and the lesion mask, leaving six nodes unassessed when required information is unavailable or evaluation is not meaningful.For example, non-mass distribution requires ductal anatomy, while the assessment category is excluded because all MAMA-MIA cases are biopsy-confirmed malignant.Source: §4.3.1 · p. 4, §4.3 · p. 4

4.3.2 Descriptor thresholds.

• Descriptor thresholds combine BI-RADS-defined and conventional kinetic cutoffs with cohort quantiles derived from expert masks and then applied unchanged to automatic masks.Quantile thresholds place lesions relative to this cohort rather than an absolute clinical criterion; the compactness threshold was calibrated to a plausible 12.8% non-mass split on expert masks, not tuned against reference labels.Source: §4.3.2 · p. 4

4.3.3 Constructing the structured BI-RADS record.

• The structured record preserves the BI-RADS sequence, including unassessed nodes, and is the language model’s only representation of a case.The MRI, mask, and continuous measurements are withheld; lesion size is compared separately.Source: §4.3.3 · p. 4, §4.3.1 · p. 4

4.4 Step 4: Descriptor-Grounded Rationale Generation

• Llama-3.3-70B verbalizes the record under instructions to describe only recorded information and provide no diagnostic verdict.It receives the decision-tree specification, record, and ten examples; a fixed template is equally faithful, while reader evaluation of the model’s more natural phrasing was not conducted.Source: §4.4 · p. 4

4. Internal enhancement

• The verbalisation model receives only the structured descriptor record, so it cannot derive new observations directly from MRI.This keeps the rationale tied to descriptors produced by the measurement and quantisation stages.Source: §4 · p. 5

4.5 Step 5: Evaluating Rationale Fidelity

• Counterfactual evaluation changes one descriptor, regenerates the rationale, and checks that the change appears without altering other fields.An assessed field is considered correct when the text says that it does not apply.Source: §4.5 · pp. 5–6
• Faithfulness, hallucination, omission, and counterfactual metrics assess how the rationale states the record, not whether the record is clinically correct.The checks cover incorrect verbalisation, unsupported content, missing information, and failure to respond to record changes.Source: §4.5 · pp. 5–6

5 IMPLEMENTATION DETAILS

• Shape descriptors use PyRadiomics 3.1.0 with B-spline interpolation; connectivity and erosion use SciPy, and volumes use SimpleITK.Ollama 0.30.6 serves the language model locally on a single NVIDIA RTX 5000 Ada GPU.Source: §5 · p. 6

6.1 Faithfulness

• 21,080 of 21,084 fields were recovered, yielding 99.98% faithfulness; hallucination and omission rates were both zero across 3,012 rationales.Descriptor extraction was completed for all 1,506 studies using expert and automatic masks.Source: §6.1 · p. 6, §6 · p. 6, Table 1 · p. 6
• Four parser errors came from one negated sentence shared across four cases; 351 of 352 distinct rationales parsed without error.The sentence expressed the correct value, but the parser did not handle negation.Source: §6.1 · p. 6

6.2 Counterfactual Consistency

• Counterfactual accuracy was 0.993 across 575 single-field perturbations; all perturbations appeared in the regenerated text, while four also changed a second parsed field.The four extra-field changes involved the same negated sentence and were parser errors.Source: §6.2 · p. 6

6.3 Descriptor Agreement under Mask Degradation

• Descriptor agreement α averaged 0.772 across 1,506 patients, versus 0.422 for shuffled pairs, and fell from 0.832 on Good masks to 0.549 on Missed masks.Table 2 reports α and mean Cohen’s κ by segmentation-quality grade; Figure 3 also plots agreement against Dice overlap and by grader.Source: §6.3 · p. 6, Table 2 · p. 7, Figure 3 · p. 7
• Delayed-kinetics agreement was 0.812 on Missed masks versus 0.938 on Good masks, while shape fell from 0.743 to 0.375 and margin from 0.766 to 0.500.Delayed kinetics averages signal over the lesion, whereas shape, margin, and internal enhancement depend on the mask boundary; delayed-kinetics κ was 0.415 on Missed masks and multiplicity had the lowest κ, 0.236.Source: §6.3 · pp. 6–7, Figure 3 · p. 7

6.4 Ablation Studies

• The shuffled-record floor reached 1.000 faithfulness, showing that faithfulness measures how well a record is stated, not whether the record is correct.A fixed template served as the ceiling reference condition.Source: §6.4 · p. 7
• Removing hand-written exemplars reduced faithfulness to 0.728 and raised omissions to 0.270; substituting raw measurements reduced faithfulness to 0.066, with omissions of 0.847 and hallucinations of 0.481.The raw-measurement condition began producing diagnostic statements that the pipeline was designed to exclude.Source: §6.4 · p. 7

7 CONCLUSION

• 99.98% faithfulness was achieved across 3,012 BI-RADS-grounded rationales generated for the 1,506 MAMA-MIA studies.The authors clarify that faithfulness measures whether rationales state their records correctly, not whether those records are clinically correct.Source: §7 · p. 7
• Descriptor agreement fell from 0.832 on Good masks to 0.549 on Missed masks.The authors intend the generated rationales as drafts for radiologist review, amendment, and sign-off.Source: §7 · p. 7

7.0.1 Limitations.

• All cases were malignant and thresholds were fitted on one dataset; descriptors remain unvalidated against expert readings and the multifocality indicator.Several measurements are simple proxies, including size-dependent surface-to-volume ratio and unnormalised heterogeneity.Source: §7.0.1 · p. 7