Intraoperative evaluation of metastatic SLNs with NIRF imaging assisted by artificial intelligence in breast cancers.

Background: Near-infrared fluorescence (NIRF) imaging with indocyanine green is widely employed for sentinel lymph node (SLN) biopsy in breast cancer but cannot assess SLN metastatic status. Current intraoperative assessment relies on frozen section (FS) analysis, which is time-consuming, causes tissue loss, and suffers from limited sensitivity with a high false-negative rate (FNR).

Materials and methods: Preclinical data included fluorescence images from 4T1-Luc and MDA-MB-231-Luc mouse lymph node metastasis models. Clinical data comprised 35 breast cancer patients in a prospective clinical trial (NCT.). Intraoperative ICG-based NIRF imaging was performed, and four convolutional neural networks (Vgg19, Efficientnet, Resnet, Densenet) were evaluated.

Results: In the mouse test set, the proposed approach achieved an area under the receiver operating characteristic curve (AUC) of 0.799 (95%CI: 0.787-0.810) for cohort 1 and 0.804 (95% CI: 0.793-0.816) for cohort 2. In the clinical cohort comprising 35 patients and 114 excised SLNs (16 metastatic, 98 non-metastatic), it demonstrated robust performance in detecting metastatic SLNs, with an AUC of 0.898 (95% CI: 0.892-0.903). The LymphNet approach, which aggregates predictions from multiple SLN images, yielded an FNR of 18.75%, comparable to FS analysis (FNR: 13.5-31.3%). LymphNet avoids tissue loss and streamlines the workflow, with the model's prediction process requiring less than 10 seconds. For the first time, we observed a marked reduction or complete absence of fluorescence in metastatic SLNs infiltrated by breast cancer cells.

Conclusion: This study establishes a novel, efficient tool for intraoperative SLN metastatic assessment.