Related Experiment Video
Updated: Aug 6, 2026
11:49
Single-port Non-liposuction Endoscopic Axillary Lymph Node Dissection in Breast Cancer Surgery
Published on: April 3, 2026
Vision Language Models for Ultrasound Assessment of Suspicious Axillary Lymph Nodes in Breast Cancer
Peng He1,2,3,4, Cong Chen5, Hai-Dong Dong6
1Department of Breast Surgery, Fujian Medical University Union Hospital, Fuzhou, Fujian Province, 350001, China.
AJR. American Journal of Roentgenology
|July 22, 2026
Summary
Vision-language models show potential in breast cancer diagnosis. GPT-5.2 demonstrated comparable accuracy to inexperienced radiologists in detecting axillary lymph node malignancy, suggesting a role as a decision-support tool.
Area of Science:
- Radiology
- Artificial Intelligence
- Oncology
Background:
- Accurate axillary lymph node characterization is crucial for individualized breast cancer management.
- The study addresses the need for precise nodal status assessment in breast cancer patients.
Purpose of the Study:
- To assess the diagnostic performance of vision-language models (VLMs) using only imaging for detecting malignancy in breast cancer axillary lymph nodes.
- To compare VLM performance against radiologists with varying experience levels.
Main Methods:
- Retrospective analysis of 718 breast cancer patients undergoing ultrasound-guided fine-needle aspiration/core-needle biopsy of axillary lymph nodes.
- Grayscale and Doppler ultrasound images of biopsied nodes were evaluated by inexperienced and experienced radiologists.
- Three VLMs (GPT-5.2, Gemini-3-Pro, Claude-Opus-4.5) analyzed images for malignancy detection.
Main Results:
- Experienced radiologists achieved the highest accuracy (83%), sensitivity (75%), and specificity (89%).
- GPT-5.2 showed accuracy (78%) comparable to inexperienced radiologists (76%) but with lower sensitivity and higher specificity.
- Inexperienced radiologists had higher sensitivity (92%) than experienced radiologists and all VLMs.
Conclusions:
- GPT-5.2, the top-performing VLM, did not significantly differ in accuracy from inexperienced radiologists.
- VLMs, particularly GPT-5.2, show promise as supervised decision-support tools to aid inexperienced radiologists and potentially reduce false positives.
