Assessing the performance of deep learning and hand-crafted radiomics models using MRI and ultrasound in predicting axillary lymph node status in breast cancer: a systematic review and meta-analysis.
Iman Kiani, Niloufar Pourakbar, Amir Mohammad Azizpour et al.
✦ AI-curated · Sources linked
In 30 seconds
This systematic review and meta-analysis assessed deep learning (DL) and hand-crafted radiomics (HCR) models using MRI or ultrasound to predict axillary lymph node metastasis in breast cancer, with histopathology as the reference standard. Across 41 studies, pooled internal-validation sensitivity was 0.79 (95% CI: 0.74-0.84) and specificity was 0.78 (95% CI: 0.75-0.81), with an AUC of 0.84. External-validation AUC was 0.82. Overall, models showed moderate diagnostic accuracy, while combined DL and HCR approaches performed better than individual methods.
Key findings
- Internal validation yielded sensitivity of 0.79 (95% CI: 0.74-0.84), specificity of 0.78 (95% CI: 0.75-0.81), and AUC 0.84.
- External validation showed sensitivity of 0.78, specificity of 0.74, and AUC of 0.82.
- Combined DL and HCR models had AUC = 0.88 in MRI and AUC = 0.92 in US, exceeding individual approaches.
- Models using intratumoral and peritumoral regions had higher AUCs (0.81 vs 0.75) than intratumoral-only models.
Why it matters
Non-invasive prediction of axillary nodal status could contribute to risk stratification and treatment planning in breast cancer. However, the pooled likelihood ratios (LR+ 3.0, LR- 0.33) indicate that these models have limited standalone clinical utility. The findings highlight the potential value of combining modelling approaches while underscoring the need for consistent methods and validation across diverse populations.
What to keep in mind
The review reports heterogeneity across studies and concludes that methodological standardization and validation across diverse populations are needed before clinical implementation.