JMIR Publications

ABSTRACT

Background:

Axillary lymph node metastasis (ALNM) serves as a critical factor in predicting breast cancer (BC). However, the noninvasive prediction of ALNM remains challenging. While some deep learning (DL) models have been developed for preoperative ALNM assessment, their performance lacks systematic evaluation.

Objective:

This study evaluates DL’s effectiveness in detecting ALNM, offering evidence to guide clinical diagnostic tools.

Methods:

Embase, Web of Science, PubMed, and Cochrane Library were searched up to May 26, 2024. The QUADAS-2 was utilized to assess the risk of bias in the selected studies. A bivariate mixed-effects model was applied, and subgroup analyses were carried out based on different imaging modalities.

Results:

21 studies (14,501 BC patients, including 4,141 LNM cases) were analyzed. Regarding ultrasound- based DL for detecting LNM, the sensitivity was 0.82 (0.77–0.86);the specificity was 0.79 (0.71–0.85); the positive likelihood ratio (LR+) was 3.9 (2.8–5.4); the negative likelihood ratio (LR-) was 0.23 (0.18–0.29); the diagnostic odds ratio (DOR) was 17 (11–25),and summary receiver operating characteristic (SROC) was 0.87 (0.26–0.99). Regarding magnetic resonance imaging (MRI)-based DL for predicting LNM, the sensitivity was 0.76 (0.68–0.83); the specificity was 0.84 (0.80–0.87); the LR+ was 4.7 (3.7–6.0); the LR- was 0.28 (0.21–0.39); the DOR was 17 (10–26); and SROC was 0.88 (0.40–0.99).

Conclusions:

Current DL approaches for detecting ALNM in BC primarily rely on US and MRI. Both US- and MRI-based DL showed good diagnostic performance, offering evidence for developing/updating clinical diagnostic models. Clinical Trial: Not applicable