acceptodds
Under review as a conference paper at ICLR 2027

A Semantically Aligned and Enriched Vision-Language Model for Breast Cancer Diagnosis on Routine Chest CT

Abstract

Routine chest CT contains the breast tissue, however, physicians often omit breast findings due to their low contrast in CT scans. In this work, we explore whether the weakly-supervised vision-language contrastive learning provides scalable opportunistic breast cancer diagnosis using routine chest CT and reports. First, a benchmark dataset is constructed (BreastCT-12K), which comprises 12,696 patients with routine chest CT scans, CT reports, and temporally matched breast MRI reports. Since CT reports often lack breast description, we curate MRI reports to retain CT-visible findings to enrich the paired CT reports. We then develop a Report-Enriched and Aligned CHest CT vision-language model for Breast cancer diagnosis (**REACH-breast**). Since whole-volume contrastive pretraining may overlook small and side-specific abnormalities of breast, REACH-breast introduces two modules to enhance fine-grained feature alignment and representation: 1) Breast-Centric Hierarchical Alignment (BCHA) matches the left, right, and bilateral breast representations with corresponding text descriptions and bridges the bilateral and global context, preserving laterality while making whole-volume features breast-aware. 2) Side-aware Evidence–Patch Association (SEPA) links report evidence to specific breast-region within patches, strengthening fine-grained visual–semantic correspondence without lesion-level annotations. Using chest CT and fixed class prompts at inference, REACH-breast achieves 89.19% AUC for zero-shot malignancy classification on the test split of BreastCT-12K, significantly exceeding the state-of-the-art breast-adapted zero-shot foundation models by at least 7.08%. It also yields the best performance in bidirectional CT-report retrieval and report generation tasks. External evaluation confirms superior performance of REACH-breast. These results demonstrate the effectiveness of vision-language modeling for opportunistic breast assessment on routine chest CT.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
3D CT, CT-CLIP, Vision-language pre-training, Breast opportunistic screening, Breast cancer diagnosis on routine chest CT

Reviews

Among:
Shape:
1234this paper, Top 30%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.67 over 3 reviews, above 70% of the others.
Reviewer_RwZjOpenReview ↗
Rating
3(+1)
Confidence
4
Informativeness
1
Presentation
3
Soundness
3
Read review
Paper summary

This paper proposes REACH-breast, a breast-aware vision–language pretraining framework for opportunistic breast cancer diagnosis using routine chest CT. It addresses two key challenges: the scarcity of breast-related descriptions in routine CT reports and the limitations of global image–text alignment in capturing fine-grained, laterality-specific breast abnormalities. Experimental results demonstrate that REACH-breast achieves 89.19% and 83.73% AUC for zero-shot breast malignancy classification on internal and external test sets, respectively. It also outperforms competing methods in bidirectional image–text retrieval, radiology report generation, and BI-RADS risk classification, demonstrating the effectiveness of report enrichment and multi-granularity semantic alignment for opportunistic breast cancer assessment on routine chest CT.

Critical strengths
  1. The authors introduce BreastCT-12K, containing 12,696 patients with paired chest CT scans, CT reports, and temporally matched breast MRI reports. Enriching CT reports with MRI-derived breast descriptions provides a practical strategy to address insufficient breast-specific language supervision.
  2. The proposed method is evaluated on multiple tasks, including zero-shot malignancy classification, BI-RADS classification, image–text retrieval, and report generation. The inclusion of an external cohort comprising 5,526 patients from three centers strengthens the empirical evaluation. Ablation studies further investigate the contributions of report enrichment and the proposed alignment modules.
Critical weaknesses
  1. The proposed BCHA module extracts breast-specific representations using anatomical masks and aligns them with corresponding textual descriptions, which is conceptually similar to anatomy-aware alignment strategies explored in prior works such as fVLM. Likewise, the maximum-similarity matching strategy in SEPA resembles existing fine-grained image–text alignment mechanisms, which has been explored in prior works such as GLoRIA. Although the authors introduce laterality-aware and cross-level alignment, the methodological distinctions from existing approaches are not sufficiently established. More direct technical comparisons and controlled experiments are needed to demonstrate the independent contributions of these designs.
  2. Insufficiently controlled comparisons on general thoracic disease recognition. Although Appendix D evaluates 16 thoracic findings on CT-RATE, Table 8 mainly relies on previously published results rather than re-evaluating competing methods under identical experimental settings. More importantly, direct comparisons with the closely related fVLM and ViSD-Boost methods are absent. The comparison with the authors' own pretrained baseline also reveals noticeable performance degradation for certain findings, such as arterial wall calcification, where AUC decreases from 88.4% to 71.1%. Therefore, the current evidence is insufficient to establish that breast-specific adaptation adequately preserves general thoracic recognition capabilities.
Questions for authors

Please refer to the weaknesses above

2 comments

Dear Authors,

I noticed that my review has not yet received a response. Could you please address the key questions and concerns raised in my review? Your response would help inform my final evaluation.

Thank you.

We sincerely thank the reviewer for the assessment and constructive questions, which have helped us improve the presentation and completeness of the manuscript.

W1. Technical distinctions and independent contributions of BCHA and SEPA.

We thank the reviewer for highlighting the connections to existing local and regional alignment approaches. The specific distinctions of our method are described below.

a. BCHA versus anatomy-aware alignment.

BCHA extracts breast-region representations from a single pass through the 3D CT encoder and introduces bidirectional cross-level alignment between whole-chest images and bilateral breast text, and between bilateral breast images and global text. Whereas fVLM establishes image–text correspondences at the same level, BCHA explicitly models bidirectional information exchange across spatial scales.

To further assess BCHA's effectiveness, we replace it with an anatomy-aware alignment strategy similar to that used in fVLM. As shown below, BCHA achieves an AUC of 88.01%, compared with 85.55% for the anatomy-aware alignment baseline, supporting the value of its laterality-aware and cross-level design.

Table r1. Internal diagnostic performance of BCHA compared with anatomy-aware alignment (%).

ModelAUCACCSpecSensF1Prec
Baseline64.5271.1890.8927.0436.6856.99
Breast CT report80.8975.4378.8267.8663.0358.85
BreastCT-12K85.2078.3581.8970.4166.7563.45
BreastCT-12K + BCHA88.0183.8688.0474.4974.0273.55
BreastCT-12K + anatomy-aware alignment85.5579.1382.1272.4568.1964.40

b. SEPA versus GLoRIA-style local alignment.

SEPA and GLoRIA share the idea of local image–text alignment, but differ in how they define local correspondences. GLoRIA primarily learns attention-based correspondences between report words and image subregions. SEPA organizes MRI descriptions into sentence-level evidence units, restricts candidate visual patches to the corresponding 3D breast region according to laterality, and computes matching scores in both the evidence-to-patch and patch-to-evidence directions. This design addresses the small size of breast lesions on chest CT and the potential for confusion between the left and right breasts.

Using the same data, backbone, and text settings, we replace SEPA with a GLoRIA-style local loss. As shown below, the GLoRIA-style loss achieves an AUC of 85.80%, compared with 86.37% for SEPA, suggesting that SEPA is better suited to this breast imaging task.

Table r2. Internal diagnostic performance of SEPA compared with a GLoRIA-style local loss (%).

ModelAUCACCSpecSensF1Prec
Baseline64.5271.1890.8927.0436.6856.99
Breast CT report80.8975.4378.8267.8663.0358.85
BreastCT-12K85.2078.3581.8970.4166.7563.45
BreastCT-12K + SEPA86.3778.4380.0774.7568.1462.61
BreastCT-12K + GLoRIA-style loss85.8077.8780.0672.9567.0562.03

W2. Preservation of general thoracic recognition under controlled comparisons.

We thank the reviewer for highlighting the scope of the comparison in Table 8. Results for CT-CLIP and its variants are obtained from the same CT-RATE test set and the same 16 findings, whereas results for BIUD and Merlin are taken from their published papers. Because the evaluation procedures are not fully standardized across methods, we treat these results as published reference values and do not interpret Table 8 as a strictly controlled ranking. fVLM and ViSD-Boost are not included in the original table because they use anatomy-conditioned inference pipelines for thoracic disease recognition; comparing their final scores directly alone cannot measure the extent to which our global representations are retained.

We agree that breast-specific adaptation does not preserve recognition performance equally across all thoracic findings. Appendix D.2 already explicitly reports this variation: AUC for arterial wall calcification decreases substantially, from 88.4% before adaptation to 71.1%. In contrast, AUC improves from 65.0% to 73.2% for lung opacity and from 86.2% to 86.7% for interlobular septal thickening. These finding-specific results indicate that the model retains useful general thoracic representations while improving breast diagnosis, although performance on some findings is substantially affected. Reducing this degradation further is a direction for our future work.

We hope these replies adequately address the concerns. We would be more than happy to conduct further analyses or provide additional clarifications if needed.

Reviewer_r6p5OpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper presents REACH-breast for opportunistic breast cancer screening on routine chest CT. The authors construct BreastCT-12K, leveraging temporally matched breast MRI reports to enrich breast-related semantics in CT, and propose BCHA and SEPA for hierarchical breast region–text alignment and fine-grained association between report evidence and CT patches. Multi-center evaluation across zero-shot vision-language tasks demonstrates competitive performance.

Critical strengths

1.This paper addresses opportunistic breast cancer screening on routine chest CT and constructs the BreastCT-12K dataset. Using temporally matched breast MRI reports to enrich limited breast-related semantics in CT offers practical value.

2.A systematic evaluation covers zero-shot diagnosis, multimodal retrieval, report generation, BI-RADS classification, and multi-center testing. Ablation studies and qualitative analyses further help validate the contributions of different modules.

Critical weaknesses

1.BCHA and SEPA are mechanistically similar to existing region-level alignment methods in medical VLP (e.g., GLoRIA, MGCA, LoVT, MedKLIP, and ASG). The work primarily extends these existing mechanisms to the breast CT setting, with limited methodological novelty.

2.The method relies on MRI-derived breast text, rarely available for routine chest CT patients. The paper neither evaluates performance without MRI supervision nor separates the contributions of laterality labels and MRI-derived lesion descriptions, leaving clinical applicability uncertain.

Questions for authors
  1. In Table 4, how much of the MRI-derived gain comes from laterality information versus lesion details (size, margins, morphology)? What is the AUC when only laterality labels are retained? This would clarify performance without MRI reports.

  2. The three external centers differ substantially in case composition (especially Center 3, where malignancy prevalence is low). Could the authors report per-center AUCs with confidence intervals to better demonstrate the stability of cross-center generalization?

  3. The model achieves a sensitivity of approximately 75%, which may be insufficient for screening applications. Could the authors report specificity at higher-sensitivity operating points (e.g., 90% and 95%) to better assess its clinical applicability?

  4. Both BCHA and SEPA rely on breast segmentation masks. Could the authors report nnU-Net’s segmentation performance (e.g., Dice) on the test set and evaluate the impact of mask quality on downstream performance?

5 comments
AuthorsOpenReview ↗

Response to Reviewer r6p5 (1/4)

We sincerely thank the reviewer for the careful and insightful assessment. The feedback has been valuable in improving the quality of our manuscript.

Q1/W2. Disentangling the gains from MRI-derived supervision.

a. Performance without MRI supervision.

We evaluate the model without MRI supervision, as reported in the second row of Table 4, which uses CT reports alone. Performance in this setting is lower than that in the third row, where MRI reports are added, demonstrating the benefit of MRI-derived supervision.

b. Contributions of laterality information.

We also conduct an experiment in which only laterality information extracted from the MRI reports is retained. Laterality information is formatted as "Focal breast finding laterality: none", "Focal breast finding laterality: left, right", "Focal breast finding laterality: right", and "Focal breast finding laterality: left". The internal test results are presented below.

Table r1. Contributions of MRI-derived laterality information and full curated MRI text (%).

ModelMRI laterality informationFull curated MRI textAUCACCSpecSensF1Prec
Baseline××64.5271.1890.8927.0436.6856.99
Breast CT report××80.8975.4378.8267.8663.0358.85
BreastCT-12K (laterality only)✓×80.9376.9382.8063.7663.0562.34
BreastCT-12K×✓85.2078.3581.8970.4166.7563.45

Abbreviations: AUC, area under the ROC curve; ACC, accuracy; Spec, specificity; Sens, sensitivity; Prec, precision.

These results demonstrate that, with CT-report supervision alone, the model achieves an AUC of 80.89%. Adding only MRI-derived laterality information raises it to 80.93%, a gain of 0.04 percentage points. Using the full curated MRI text raises AUC to 85.20%, a further gain of 4.27 percentage points over the laterality-only setting. This comparison suggests that the improvement comes chiefly from information beyond laterality, including lesion descriptions of size, margins, and morphology.

AuthorsOpenReview ↗

Response to Reviewer r6p5 (2/4)

Q2. Results of the three external centers.

a. Differences in case composition across centers.

The three external datasets are collected from different hospitals and therefore differ substantially in case composition. Center 3 contains a small proportion of malignant cases to approximate a setting in which breast malignancy is uncommon, reflecting the low prevalence encountered in practice.

b. Per-center performance and AUC confidence intervals.

The results for each center and per-center AUCs with 95% confidence intervals are reported below. Our model achieves AUCs of 75.32% (95% CI: 72.43–78.07), 84.84% (82.51–86.97), and 89.35% (85.27–92.93) at Centers 1, 2, and 3, respectively, with the highest AUC point estimate among the evaluated methods at every center. Although absolute performance varies across centers, this consistent advantage supports cross-center generalization, including at Center 3, where malignancy prevalence is low.

Table r2. AUCs with 95% confidence intervals for each external center and the macro-average across centers (%).

MethodCenter 1 (n = 1,284)Center 2 (n = 1,186)Center 3 (n = 3,053)Macro-average across centers
Base model62.85 (59.34–66.19)44.89 (41.55–48.20)84.54 (81.25–88.79)64.09 (62.25–66.24)
CT-LiPro58.42 (54.82–62.02)40.50 (37.16–43.90)62.16 (57.43–67.67)53.69 (51.48–56.21)
CT-VocabFine53.95 (50.22–57.59)38.31 (34.99–41.68)64.75 (60.38–69.90)52.34 (50.17–54.84)
CT-CLIP61.56 (57.88–65.06)25.05 (22.08–28.10)63.22 (58.07–69.39)49.94 (47.68–52.56)
Merlin62.59 (59.01–66.04)53.88 (50.35–57.35)68.10 (63.21–73.56)61.52 (59.21–63.92)
fVLM40.55 (37.05–44.16)83.24 (80.64–85.75)76.76 (72.14–82.60)66.85 (64.83–69.32)
ViSD-Boost47.47 (43.95–51.10)79.44 (76.80–82.04)66.83 (61.33–73.23)64.58 (62.23–67.22)
Ours75.32 (72.43–78.07)84.84 (82.51–86.97)89.35 (85.27–92.93)83.17 (81.35–84.86)

Table r3. Diagnostic performance at Center 1 on the shared cohort of 1,284 patients (954 malignant and 330 non-malignant) (%).

MethodAUCACCF1Prec
Base model62.8545.6446.6486.16
CT-LiPro58.4229.1311.9977.50
CT-VocabFine53.9526.874.4875.86
CT-CLIP61.5648.3652.3483.30
Merlin62.5938.6334.9882.17
fVLM40.5574.3085.2374.38
ViSD-Boost47.4766.2078.6674.07
Ours75.3267.2975.4785.22

Table r4. Diagnostic performance at Center 2 on the shared cohort of 1,186 patients (790 malignant and 396 non-malignant) (%).

MethodAUCACCF1Prec
Base model44.8944.6042.7268.63
CT-LiPro40.5035.5024.3355.66
CT-VocabFine38.3134.4015.6254.55
CT-CLIP25.0530.7838.0447.10
Merlin53.8841.1529.4973.00
fVLM83.2467.7978.8770.04
ViSD-Boost79.4468.5571.2091.29
Ours84.8472.6876.1491.02

Table r5. Diagnostic performance at Center 3 on the shared cohort of 3,053 patients (110 malignant and 2,943 non-malignant) (%).

MethodAUCACCF1Prec
Base model84.5489.5828.7019.05
CT-LiPro62.1673.2110.505.97
CT-VocabFine64.7588.149.056.25
CT-CLIP63.2265.9410.966.05
Merlin68.1094.1417.5117.76
fVLM76.7635.908.934.71
ViSD-Boost66.8394.9928.8429.52
Ours89.3593.9145.6133.62
AuthorsOpenReview ↗

Response to Reviewer r6p5 (3/4)

Q3. Specificity at high sensitivity.

We thank the reviewer for drawing attention to specificity at high sensitivity. We evaluate these operating points on the test set of 1,270 patients (392 malignant and 878 non-malignant). The results are presented below.

Table r6. Specificity at high sensitivity operating points on the internal test cohort of 1,270 patients.

Sensitivity74.75%80.10%85.20%90.05%93.10%95.15%
Specificity91.80%83.94%77.45%60.14%48.30%40.77%

Our study targets opportunistic screening in patients who have already undergone chest CT. The model is intended to flag suspicious cases for further assessment, rather than provide a definitive diagnosis of breast cancer. At sensitivities of 90.05% and 95.15%, specificity is 60.14% and 40.77%, respectively. At a sensitivity of 85.20%, the model retains a specificity of 77.45%, providing an additional reference point alongside the higher-sensitivity settings.

These results quantify the trade-off between detecting more malignant cases and increasing false positives and the associated follow-up burden, providing evidence to inform operating-point selection for different screening priorities. We will include the complete results in the final manuscript to make this trade-off explicit.

Q4. Breast mask quality and its impact.

We train our own nnU-Net model for breast ROI segmentation using 500 physician-annotated CT scans from another hospital, rather than using off-the-shelf breast segmentation weights. The model achieves a mean breast ROI Dice score of 97.18% in five-fold out-of-fold validation.

We then use the trained weights to generate breast masks for the data in this study, providing regional guidance for BCHA and SEPA during training. Zero-shot diagnostic inference uses only CT images and does not require breast masks; consequently, inference does not depend on the segmentation model. There is also precedent for using automatically generated anatomical masks for image–text alignment, for example, fVLM uses the nnU-Net-based TotalSegmentator to obtain organ masks.

Reviewer_KsBMOpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

REACH-breast aims to assess breast disease from routine chest CT. The authors enrich CT reports with descriptions from MRI exams taken around the same time. The model learns to connect report information with the whole image and with breast regions. The experiments cover diagnosis, retrieval, report generation, and breast risk classification. At diagnosis time, the model uses CT images and fixed prompts; it does not need the patient's MRI report.

Critical strengths
  • The paper addresses a practical problem: routine CT reports may leave out useful breast findings. Adding information from MRI reports is a sensible way to improve training. The model shows clear gains in diagnostic ranking over the listed breast-adapted baselines on internal and external data. The external tests and the stated separation of patients between data splits strengthen the evidence.
  • The component studies support the main design choices. They show benefits from adapting to CT reports, adding MRI information, and using the full model. Tests of the training objectives and improvements on retrieval and other tasks provide further support. The appendix also discusses limitations and errors in generated reports openly.
Critical weaknesses
  • It is not clear whether every method is evaluated on the same patients. The appendix allows methods that need segmentation masks, including some component tests, to omit cases. The authors should report which cases were included and compare the methods on a shared set of patients. Results from each external center would also help. Choosing decision thresholds on the test data limits what the reported classification performance says about real use, although it does not remove the main evidence from AUC.
  • The clinical claims need care. The added MRI descriptions are checked against source reports, but it is unclear how often the findings are visible on the paired CT. The report-generation evaluation mainly measures wording overlap, and the examples contain clinically relevant errors. The association maps also do not establish precise lesion localization. These limitations affect the proposed clinical uses, while the gains in learned image features remain useful.
Questions for authors
  • Which positive and negative cases were used for each method and component test? Does the diagnostic advantage remain when every method is evaluated on the same patients, and when external centers are considered separately?
  • Did you check whether the MRI findings used for training were visible on the paired CT images? Please describe that check and clarify which claims are supported about report accuracy and lesion localization.
2 comments
AuthorsOpenReview ↗

Response to Reviewer KsBM (1/2)

We sincerely thank the reviewer for the careful and constructive feedback. Our point-by-point responses are provided below.

Q1/W1. Fair diagnostic evaluation on shared patients and across external centers.

a. Evaluation cohorts and comparison on shared patients.

For the internal results in Table 1, fVLM and ViSD-Boost are evaluated on 1,267 cases (392 malignant, 875 non-malignant). All other methods are evaluated on 1,270 cases (392 malignant, 878 non-malignant).

For external datasets, at Center 1, fVLM is evaluated on 1,284 cases (954 malignant, 330 non-malignant), ViSD-Boost on 1,286 cases (956 malignant, 330 non-malignant), and all other methods on 1,287 cases (957 malignant, 330 non-malignant). At Center 2, all methods are evaluated on 1,186 cases (790 malignant, 396 non-malignant). At Center 3, all methods are evaluated on 3,053 cases (110 malignant, 2,943 non-malignant).

To ensure a fair comparison, we re-evaluate all methods on the same patients. The shared internal test cohort comprises 1,256 patients: 392 malignant and 864 non-malignant. The shared external cohorts comprise 1,284 patients at Center 1 (954 malignant and 330 non-malignant), 1,186 patients at Center 2 (790 malignant and 396 non-malignant), and 3,053 patients at Center 3 (110 malignant and 2,943 non-malignant). The results are presented below.

Table r1. Diagnostic performance on the shared internal test cohort and macro-averaged performance across the three external centers (%).

MethodInternal AUCInternal ACCInternal F1Internal PrecExternal AUCExternal ACCExternal F1External Prec
Base model79.8978.2652.3682.8764.0959.9439.3557.94
CT-LiPro80.5876.1960.5062.7453.6945.9415.6146.38
CT-VocabFine82.4377.3955.0672.5052.3449.809.7145.55
CT-CLIP75.7470.7854.4153.0349.9448.3633.7845.48
Merlin80.8474.5237.5080.0061.5257.9727.3357.64
fVLM67.7661.6252.3742.7466.8559.3357.6849.71
ViSD-Boost81.8576.2746.0179.3864.5876.5859.5764.96
Ours89.1686.3977.4180.2783.1777.9665.7469.96

b. Results for individual external centers.

The results for each external center are presented below, demonstrating the advantages of our method.

Table r2. Diagnostic performance at Center 1 on 1,284 patients (954 malignant, 330 non-malignant) (%).

MethodAUCACCF1Prec
Base model62.8545.6446.6486.16
CT-LiPro58.4229.1311.9977.50
CT-VocabFine53.9526.874.4875.86
CT-CLIP61.5648.3652.3483.30
Merlin62.5938.6334.9882.17
fVLM40.5574.3085.2374.38
ViSD-Boost47.4766.2078.6674.07
Ours75.3267.2975.4785.22

Table r3. Diagnostic performance at Center 2 on 1,186 patients (790 malignant, 396 non-malignant) (%).

MethodAUCACCF1Prec
Base model44.8944.6042.7268.63
CT-LiPro40.5035.5024.3355.66
CT-VocabFine38.3134.4015.6254.55
CT-CLIP25.0530.7838.0447.10
Merlin53.8841.1529.4973.00
fVLM83.2467.7978.8770.04
ViSD-Boost79.4468.5571.2091.29
Ours84.8472.6876.1491.02

Table r4. Diagnostic performance at Center 3 on 3,053 patients (110 malignant, 2,943 non-malignant) (%).

MethodAUCACCF1Prec
Base model84.5489.5828.7019.05
CT-LiPro62.1673.2110.505.97
CT-VocabFine64.7588.149.056.25
CT-CLIP63.2265.9410.966.05
Merlin68.1094.1417.5117.76
fVLM76.7635.908.934.71
ViSD-Boost66.8394.9928.8429.52
Ours89.3593.9145.6133.62

c. Updates to the manuscript. We will revise Tables 1, 6, and Figure 3 to ensure all methods are evaluated on the same internal cohort of 1,256 patients (392 malignant and 864 non-malignant). We will also revise Table 5 to report the shared patient cohort at Center 1 and correct the previously misstated numbers of malignant and non-malignant cases at Center 2.

d. Threshold selection. We agree that selecting thresholds on test data limits the informativeness of threshold-dependent metrics for real-world deployment. Following the CT-CLIP convention, we determine an operating threshold for each method using only the internal test set and apply it unchanged to all three external centers, without center-specific adjustment. We will describe this protocol more explicitly in the revised manuscript. The tables above report results for each external center and their macro-average, demonstrating cross-center robustness under fixed thresholds.

AuthorsOpenReview ↗

Response to Reviewer KsBM (2/2)

Q2/W2. CT visibility of MRI-derived findings and limits of clinical claims.

We thank the reviewer for this question. We use curated MRI-derived descriptions rather than complete MRI reports as supervision. Under clinical guidance, we remove descriptions of MRI-specific signal characteristics, enhancement, and ducts, while retaining information that could potentially correspond to CT appearances, such as lesion laterality, location, approximate size, and margins. This approach does not require every MRI finding to have already been recognized by a radiologist reading the routine CT. Instead, paired MRI reports provide breast-related semantic information that may be absent from routine CT reports, allowing us to explore whether the model can learn associated subtle imaging patterns from CT.

Paired MRI reports provide breast-related information often absent from routine CT reports. The ablation in Table 4 and the diagnostic AUC gains in Table 1 support the value of this supervision for learning CT representations. Figure 6 provides supporting evidence: nodule descriptions are primarily associated with image patches in one breast, whereas density descriptions are associated with bilateral regions. The report-generation results indicate that the model includes breast-related content more frequently, but the current lexical-overlap metrics and examples are insufficient to establish the clinical accuracy of the generated reports. We will clarify the scope of this evidence in the revised manuscript.

We hope these responses address the reviewers’ concerns. We would be glad to run further analyses if any questions remain.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.