Ultrasound derived deep learning features for predicting axillary lymph node metastasis in breast cancer using graph...

Area of Science:

• Medical Imaging
• Artificial Intelligence in Oncology
• Breast Cancer Diagnostics

Background:

• Axillary lymph node metastasis (ALNM) is a critical prognostic factor in breast cancer.
• Accurate preoperative assessment of ALNM is essential for treatment planning and avoiding overtreatment.
• Current diagnostic methods for ALNM can be invasive and may not always be definitive.

Purpose of the Study:

• To develop and validate an ultrasound-based graph convolutional network (US-based GCN) model for predicting ALNM in breast cancer patients.
• To assess the performance of the US-based GCN model in independent validation cohorts.
• To explore the potential of the US-based GCN model in guiding ALNM management and preventing overtreatment.

Main Methods:

• Retrospective enrollment of 820 breast cancer patients who underwent preoperative ultrasonography (US).
• Development of a US-based GCN model utilizing deep learning features from US images.
• Validation of the model using two independent cohorts (112 and 87 patients).

Main Results:

• The US-based GCN model demonstrated satisfactory performance in validation cohort 1 (AUC: 0.88, accuracy: 0.76).
• The model also performed satisfactorily in validation cohort 2 (AUC: 0.84, accuracy: 0.75).
• The developed model shows potential for noninvasive ALNM detection and clinical decision support.

Conclusions:

• A US-based GCN model was successfully developed and validated for assessing axillary lymph node status in breast cancer patients preoperatively.
• This AI-driven approach offers a promising noninvasive method for detecting ALNM, aiding clinical decision-making.
• Prospective studies are anticipated to provide high-level evidence for the clinical utility of this US-based GCN model.