Point-of-Care Ultrasound in Diagnosis of Breast Cancer-Related Lymphedema: A Multimodal Database and Deep Learning Approach.
Authors
Abstract
Breast Cancer-Related Lymphedema (BCRL) is a common complication after breast cancer treatment, leading to chronic arm swelling and reduced quality of life. Early detection remains challenging due to limitations in current assessment methods and a shortage of expert clinicians. Point-of-care ultrasound (POCUS) imaging offers a cost-effective, non-invasive tool for BCRL evaluation; however, the visual disparity between affected and unaffected arms can be subtle. To address this, we introduce a multimodal dataset containing B-mode images, radio-frequency (RF) data, perometry-based arm volume measurements, the number of dissected lymph nodes, the time elapsed since surgery, and body mass index (BMI) from 57 subjects. Additionally, the dataset includes an assessment of the severity of BCRL by an expert clinician (A.T.). We also propose a deep learning-based framework for automated BCRL diagnosis using multimodal data. Our method first employs a Siamese-based architecture to extract and compare features from bilateral arm images, addressing subject-specific variability. We further improve the method by incorporating clinical tabular data and investigating the effect of multimodal fusion and supervised contrastive learning on the model's performance. To our knowledge, this is the first automated deep learning approach for BCRL diagnosis. Using three-fold cross-validation, the proposed Multimodal model achieved an AUC of 91.0%, an accuracy of 87.0%, and an F1 score of 83.5%. It outperformed the two other variants, Unimodal-A and Unimodal-B, in BCRL classification across most metrics. The source code and trained model will be released at http://code.sonography.ai.