Source-linked AI summary
TRACE: Training-time Report-guided and Clinically Ordered Concept Editing
Wentao Yue, Tianyou Lai, Jiayu Luo, Qingyu Mao, Ziying Wang, Zhenyuan Ning, Qilei Li
TL;DR
Breast ultrasound models need clinically grounded concepts but often require complete annotations or reports at test time, limiting practical image-only use. TRACE uses structured reports as privileged training supervision to edit image-derived concepts, and experiments report superior performance with cross-domain robustness while retaining image-only deployment.
Problem
Existing breast ultrasound methods often lack clinically grounded intermediate supervision, assume complete concept annotations, or require reports during inference.
Method
TRACE uses report-derived concepts during training to guide editing of coarse image concepts in a malignancy-aware ordered space, with SCMT and an image-only self-editor distilled from the teacher.
Results
TRACE achieves superior in-domain performance and strong cross-domain robustness across BUSC and external datasets compared with existing methods.
Takeaways & Limitations
Learning concept correction from training-only structured supervision supports clinically meaningful, generalizable breast ultrasound diagnosis without reports at test time.
Takeaways & Limitations
Cross-domain deployment assumes structured reports and concept annotations are available during training but absent in the target domain, and only a subset of training samples has expert concept vectors.
Abstract
from arXiv · showhide
Breast ultrasound diagnosis relies on clinically meaningful semantic concepts, yet most deep learning methods adopt end-to-end image-to-label paradigms that lack interpretability and robustness. While concept-based approaches offer a promising alternative, they often assume complete annotations or require multimodal inputs at inference, which significantly limits their real-world applicability. To tackle these issues, we propose Training-time Report-guided and Clinically Ordered Concept Editing (TRACE), a training-time report-guided framework that leverages structured radiology reports as privileged concept supervision while enabling image-only diagnosis at test time. TRACE refines image-derived concepts through a teacher-guided editing mechanism within a malignancy-aware ordered concept space. To address incomplete annotations, we introduce Strategic Concept Missing Training (SCMT) and train an image-only self-editor via edit distillation for autonomous concept refinement. Besides, we introduce BUSC, a concept-enriched benchmark linking images, labels, and structured attributes. Experiments across multiple datasets demonstrate that TRACE achieves superior performance and improved cross-domain robustness compared to existing methods.
1 Introduction
Breast ultrasound diagnosis needs clinically grounded concepts, but existing methods often lack structured supervision, require complete annotations, or depend on reports at inference. TRACE uses reports during training to refine image-derived concepts while supporting image-only diagnosis and robustness under incomplete concepts.
- End-to-end image-to-label models can rely on shortcut visual correlations, limiting interpretability and robustness under distribution shifts or incomplete semantic supervision.
- Concept-based methods improve clinical alignment but commonly assume complete concept supervision or require text during inference.
- BUSC links breast ultrasound images, diagnostic labels, and structured clinical concepts across the BUSC-BUSI647 and BUSC-BUSBRA development subsets.
- TRACE treats structured reports as privileged training-only concept supervision, using a teacher-guided editor to revise coarse image-derived concepts within a malignancy-aware ordered space.
- SCMT masks report concepts according to clinically informed missing patterns, while edit distillation trains an image-only self-editor for test-time concept refinement.
- TRACE consistently achieves superior in-domain performance and strong cross-domain robustness across BUSC and three external datasets compared with diverse baseline families.
2 Related Work
Prior breast ultrasound methods incorporate structured concepts or privileged information, but commonly require concepts, reports, or complete annotations during evaluation. TRACE addresses this setting by learning from training-time structured supervision while targeting image-only deployment.
- Structured BI-RADS-related attributes such as shape, margin, orientation, posterior acoustic features, and echogenicity support breast ultrasound lesion assessment and report writing.
- Most existing structured-attribute methods assume concepts or reports are available during evaluation, or require complete concept annotations within the same domain.
- Recent multimodal approaches fuse radiology reports with images but still require reports as inference-time inputs.
- Interpretable models such as Explicd, VLG-CBM, and MVP-CBM learn concept representations through textual alignment, grounded detection, or preference modeling.
- Medical privileged-information methods typically use extra information for distillation, prototype constraints, or auxiliary modalities rather than explicitly modeling structured breast ultrasound concept editing.
3 Methodology
TRACE uses structured concepts as training-only teacher supervision to edit coarse image-derived concepts in a clinically ordered space, while preserving image-only diagnosis at deployment. It addresses incomplete supervision through hierarchical masking and distills teacher-guided editing into a self-editor.
- Problem setting: TRACE learns image-only diagnosis from limited structured concept supervision, with reports available during training but removed at inference.The test setting provides images and labels without structured concepts or report information.
- Report-guided concept editing: An image encoder and initial concept predictor produce coarse concepts, which a report-guided editor corrects using teacher residuals rather than predicting concepts from scratch.The teacher concept specifies the desired semantic state, while the residual represents the correction direction and magnitude.
- Clinically ordered concept space: Clinically ordered constraints make editing directions consistent with malignancy-related risk order and reduce medically inconsistent label transitions.The ordering applies to concepts such as margin and shape, which correspond to different malignancy-risk levels.
- Hierarchical concept missing training: Strategic Concept Missing Training hierarchically masks concepts, preserving key attributes more often while making peripheral attributes more likely to be missing.The hierarchy uses key, mid-level, and peripheral concept sets with preservation probabilities r1 = 0.2, r2 = 0.5, and r3 = 0.8 as specified in the training design.
- Hierarchical concept missing training: The masked editing objective trains cross-concept compensation, narrowing the mismatch between report-supervised training and report-free deployment.The design forces dependency and redundancy compensation under incomplete concept supervision.
- Inference and training objectives: At test time, an image-only self-editor transfers the privileged teacher’s correction behavior to refine concepts and produce the final diagnosis.Additional consistency training reduces the mismatch between teacher-guided editing during training and self-editing during inference.
4 Experiments
TRACE is evaluated on in-domain BUSC subsets and zero-shot external datasets under image-only inference, with ablations examining editors, missing concepts, supervision scarcity, and feature representations. It achieves the strongest in-domain results and robust cross-domain performance, while attention editing and partial supervision remain effective.
- Experimental Setup: TRACE is evaluated on BUSC-BUSBRA, BUSC-BUSI647, and three external datasets for zero-shot cross-domain generalization.External evaluation removes structured reports and concept annotations from the target data, reflecting image-only deployment.
- Experimental Setup: All methods use a unified image-only test protocol, with structured reports and concept labels available only during training.The comparison includes standard visual backbones, concept bottleneck models, prototype-based methods, explainable classifiers, and vision-language models.
- Main Results: TRACE achieves the best AUC, Accuracy, and F1 on both BUSC-BUSBRA and BUSC-BUSI647, outperforming image-only and concept-based baselines.On BUSC-BUSBRA, the comparison includes ResNet50, PCBM, MVP-CBM, and Explicd; TRACE also remains strongest on all three metrics for BUSC-BUSI647.
- Cross-Domain Generalization: TRACE achieves the best Accuracy on Ardakani, the best AUC on BUS_UC, and the best AUC and Accuracy on BrEaST.It also achieves the best Accuracy and one of the strongest AUC results when transferring to BUSC-BUSI647∗.
- Ablation Study: Attention editing provides the most reliable overall cross-domain choice, although MLP and gating show dataset- or metric-specific strengths.TRACE-Att achieves the best or tied-best results across multiple target domains, while MLP has higher AUC on some domains and gating remains competitive in Accuracy.
- Ablation Study: Missing-concept training improves several BUSC-BUSBRA metrics, whereas Full training performs better on BUSC-BUSI647, showing dataset-dependent effects.TRACE remains effective when supervision is reduced from 100% to 50% and 20%; attention and gating show more stable trends, while MLP has a higher full-supervision upper bound.
- Representation Analysis: TRACE embeddings show clearer benign–malignant separation, with more compact intra-class clusters and less cross-class overlap than competing concept models.The representation-level pattern is consistent with the reported AUC, Accuracy, and F1 improvements.
5 Conclusion
TRACE uses training-only structured reports to refine coarse image-derived concepts while enabling image-only breast ultrasound diagnosis at test time. Across in-domain and zero-shot cross-domain evaluations, it outperforms strong baselines and supports robust deployment across datasets.
- TRACE treats structured reports as privileged concept teachers during training rather than inference-time inputs.The framework learns how coarse image-derived concepts should be refined using report supervision.
- TRACE combines teacher-guided concept editing, a clinically ordered concept space, Strategic Concept Missing Training, and an image-only self-editor.These components target concept refinement under incomplete concepts while preserving image-only inference.
- TRACE introduces BUSC, a benchmark aligning ultrasound images, diagnostic labels, and structured clinical concepts.BUSC supports concept-enriched evaluation for breast ultrasound diagnosis.
- TRACE outperforms strong baselines and supports robust image-only deployment across datasets.The reported experiments include both in-domain and zero-shot cross-domain settings.