Construction of a Benchmark for Breast Ultrasound AI Interpretation and Performance Evaluation of Multimodal AI Models

Official title

Construction of a Standardized Benchmark Evaluation System for Intelligent Breast Ultrasound Image Interpretation and Systematic Performance Assessment of Multimodal Artificial Intelligence Models Based on ACR BI-RADS v2025 Criteria

Overview

This single-center, retrospective, observational study aims to construct a standardized benchmark evaluation system for intelligent breast ultrasound image interpretation and to systematically assess the diagnostic performance of current mainstream multimodal artificial intelligence (AI) models. De-identified B-mode breast ultrasound images with confirmed pathological diagnoses will be retrospectively collected from the institutional archive (2018-2025) and supplemented with images from published open-access datasets. Expert radiologists with varying experience levels will independently annotate all images according to the American College of Radiology (ACR) Breast Imaging Reporting and Data System (BI-RADS) v2025 criteria, including glandular tissue composition, lesion characterization (mass vs. non-mass lesion), morphological descriptors, and final BI-RADS classification. Baseline deep learning models (CNN-based ResNet-50 and Transformer-based USFM) will be trained to establish performance baselines and to stratify cases by diagnostic difficulty through cross-architecture consensus. Multiple multimodal large language models (MLLMs), including both general-purpose and medical-domain models, will then be evaluated via standardized API calls using BI-RADS-guided chain-of-thought prompts at temperature 0 for reproducibility. Primary endpoints include BI-RADS classification accuracy and diagnostic AUC for benign-malignant differentiation. Model robustness and safety will be assessed through out-of-distribution rejection testing, temperature-stability experiments, and thinking-mode ablation studies. This study adheres to the FLAIR and TRIPOD-LLM reporting guidelines.

Detailed description

Background: Breast cancer is the most prevalent malignancy among women worldwide. Ultrasound is a first-line screening modality, particularly in Asian populations with dense breast tissue where mammographic sensitivity is limited. However, ultrasound interpretation is highly operator-dependent, with substantial inter-observer variability in BI-RADS classification, especially for category 4A-4B lesions. Multimodal large language models (MLLMs) have emerged as a promising tool for medical image analysis due to their zero-shot diagnostic capability, interpretable chain-of-thought reasoning, and structured report generation. Nevertheless, there is currently no standardized benchmark for evaluating AI performance in breast ultrasound interpretation.

Study Design: Approximately 1,380 breast ultrasound images will be curated (1,200 evaluation set + 150 out-of-distribution safety test set + 30 prompt development set), encompassing three diagnostic categories: normal breast, benign lesions (BI-RADS 2-4B), and malignant lesions (BI-RADS 3-5). Two junior radiologists (\<5 years of experience) and two senior radiologists (\>15 years) will independently annotate images per ACR BI-RADS v2025 with arbitration by a fifth expert for discordant cases.

Diagnostic difficulty will be stratified into three tiers using cross-architecture deep learning consensus: Tier 1 (straightforward, both models correct), Tier 2 (equivocal, one correct/one incorrect), and Tier 3 (difficult, both incorrect, with senior expert validation). MLLMs will be evaluated across multiple dimensions: classification accuracy, sensitivity, specificity, F1 score, AUC, Cohen's kappa agreement with expert consensus, expected calibration error (ECE), morphological feature description accuracy, and chain-of-thought reasoning quality.

Safety Assessment: (1) Out-of-distribution rejection test using 150 non-diagnostic images (degraded images, non-breast ultrasound, other imaging modalities); (2) Temperature-stability pre-experiment across parameter settings; (3) Thinking-mode ablation comparing standard vs. chain-of-thought reasoning modes. All experiments use fixed model snapshots, system fingerprint monitoring, and complete logging for reproducibility.

Interventions

• Diagnostic test Multimodal AI Model Diagnostic Evaluation
  Retrospective evaluation of de-identified breast ultrasound images by multiple AI systems, including baseline deep learning models (ResNet-50, USFM) and multimodal large language models, using standardized BI-RADS-guided chain-of-thought prompts via API. No patient contact or clinical decision-making is involved.

Primary outcome measures

• Diagnostic Accuracy for Pathological Diagnosis [Time frame: At study completion, approximately 12 months]
• BI-RADS Classification Accuracy [Time frame: At study completion, approximately 12 months]

Secondary outcome measures (3)

• Agreement with Expert Consensus (Cohen's Kappa) [Time frame: At study completion, approximately 12 months]
• Out-of-Distribution Rejection Rate [Time frame: At study completion, approximately 12 months]
• Sensitivity, Specificity, PPV, NPV, and F1 Score [Time frame: At study completion, approximately 12 months]

Eligibility criteria

Inclusion criteria

• B-mode breast ultrasound grayscale images from the institutional PACS database or from published open-access breast ultrasound datasets with documented original institutional ethics approval
• Image quality adequate for clinical diagnosis with clear visualization of the region of interest
• Pathological diagnosis confirmed (for benign and malignant lesion groups), or normal breast status confirmed by a senior radiologist with >15 years of breast ultrasound experience (for the normal group)
• Complete de-identification with removal of all personally identifiable information

Exclusion criteria

• Severely degraded image quality precluding meaningful BI-RADS assessment
• Duplicate images from the same patient (only the most representative image retained per lesion)
• Images with residual personally identifiable information after de-identification processing
• Cases with ambiguous, disputed, or unavailable pathological results
• Non-B-mode ultrasound images, including elastography, contrast-enhanced ultrasound, and Doppler imaging

Criteria are shown verbatim from the registry (in English). Final eligibility is always assessed by the study center.

Healthy volunteers: Yes

Study design

Observational model

Cohort

Study locations

China · 1 center

• Peking Union Medical College Hospital — Beijing

Publications

• Bi WL, Hosny A, Schabath MB, Giger ML, Birkbak NJ, Mehrtash A, Allison T, Arnaout O, Abbosh C, Dunn IF, Mak RH, Tamimi RM, Tempany CM, Swanton C, Hoffmann U, Schwartz LH, Gillies RJ, Huang RY, Aerts HJWL. Artificial intelligence in cancer imaging: Clinical challenges and applications. CA Cancer J Clin. 2019 Mar;69(2):127-157. doi: 10.3322/caac.21552. Epub 2019 Feb 5. PMID 30720861
• Bhayana R, Krishna S, Bleakney RR. Performance of ChatGPT on a Radiology Board-style Examination: Insights into Current Strengths and Limitations. Radiology. 2023 Jun;307(5):e230582. doi: 10.1148/radiol.230582. Epub 2023 May 16. PMID 37191485
• Clusmann J, Kolbinger FR, Muti HS, Carrero ZI, Eckardt JN, Laleh NG, Loffler CML, Schwarzkopf SC, Unger M, Veldhuizen GP, Wagner SJ, Kather JN. The future landscape of large language models in medicine. Commun Med (Lond). 2023 Oct 10;3(1):141. doi: 10.1038/s43856-023-00370-1. PMID 37816837
• Seyyed-Kalantari L, Zhang H, McDermott MBA, Chen IY, Ghassemi M. Underdiagnosis bias of artificial intelligence algorithms applied to chest radiographs in under-served patient populations. Nat Med. 2021 Dec;27(12):2176-2182. doi: 10.1038/s41591-021-01595-0. Epub 2021 Dec 10. PMID 34893776
• Moor M, Banerjee O, Abad ZSH, Krumholz HM, Leskovec J, Topol EJ, Rajpurkar P. Foundation models for generalist medical artificial intelligence. Nature. 2023 Apr;616(7956):259-265. doi: 10.1038/s41586-023-05881-4. Epub 2023 Apr 12. PMID 37045921
• Sung H, Ferlay J, Siegel RL, Laversanne M, Soerjomataram I, Jemal A, Bray F. Global Cancer Statistics 2020: GLOBOCAN Estimates of Incidence and Mortality Worldwide for 36 Cancers in 185 Countries. CA Cancer J Clin. 2021 May;71(3):209-249. doi: 10.3322/caac.21660. Epub 2021 Feb 4. PMID 33538338
• Benary M, Wang XD, Schmidt M, Soll D, Hilfenhaus G, Nassir M, Sigler C, Knodler M, Keller U, Beule D, Keilholz U, Leser U, Rieke DT. Leveraging Large Language Models for Decision Support in Personalized Oncology. JAMA Netw Open. 2023 Nov 1;6(11):e2343689. doi: 10.1001/jamanetworkopen.2023.43689. PMID 37976064
• Miaojiao S, Xia L, Xian Tao Z, Zhi Liang H, Sheng C, Songsong W. Using a Large Language Model for Breast Imaging Reporting and Data System Classification and Malignancy Prediction to Enhance Breast Ultrasound Diagnosis: Retrospective Study. JMIR Med Inform. 2025 Jun 11;13:e70924. doi: 10.2196/70924. PMID 40498674

Identifiers

NCT: NCT07500428 · K10349 · 2024-I2M-CT-B-035 · I-26PJ0568

Primary sources (government registries)

• ClinicalTrials.gov (US NIH)

View this study on ClinicalTrials.gov ↗