Table 3 Mammograms largely enhanced the performance metrics for the prediction of lymph node metastasis in the independent test set at a sensitivity ≥90%

From: Deep learning on routine full-breast mammograms enhances lymph node metastasis prediction in early breast cancer

Metrics

PreopClinic

FullMammo

PreopClinic + FullMammo

PreopClinic + Tsize&Multifoc

ROC AUC

0.706 (±0.055)

0.782 (±0.047)

0.782 (±0.050)

0.740 (±0.051)

PR AUC

0.468 (±0.073)

0.522 (±0.077)

0.504 (±0.077)

0.461 (±0.075)

Sensitivity (recall), %

91.0 (±2.6)

91.0 (±2.7)

91.2 (±2.9)

91.1 (±2.7)

Specificity (TNR), %

31.9 (±13.5)

50.9 (±12.3)

50.3 (±16.4)

45.7 (±15.7)

PPV (precision), %

28.9 (±4.5)

36.4 (±6.5)

36.8 (±8.6)

34.4 (±7.0)

NPV, %

91.5 (±3.9)

94.9 (±1.7)

94.9 (±2.2)

94.2 (±2.4)

Accuracy, %

45.4 (±10.2)

60.0 (±9.3)

59.6 (±12.4)

56.0 (±11.9)

SLNB reduction rate, %

27.0 (±11.0)

41.1 (±10.2)

41.7 (±13.0)

37.3 (±12.1)

Net benefit, %

3.6 (±1.7)

12.8 (±2.0)

10.9 (±2.2)

9.0 (±1.8)

  1. Performance metrics were calculated using the same independent test set as previous (site 2, N = 123, LNM positive rate = 22.8%), with the addition of 14 patients for whom ROI annotations were unavailable in Fig. 3 and Table 2. Mean and standard deviation were calculated across 1000 bootstrap samples. The best mean value among all models is denoted in bold. Net benefit measures the trade-off between benefit (true positives) and harm (false positives) at the threshold when sensitivity is no less than 90%. Decision curve analysis of net benefit against [0,1] threshold is provided in Supplementary Fig. 4.
  2. PreopClinic preoperative clinicopathology, fullMammo full-breast mammogram, Tsize tumor size, Multifoc multifocality, ROC receiver operating characteristics, AUC area under the curve, PR precision recall, TNR true negative rate, PPV positive predictive value, NPV negative predictive value, SLNB sentinel lymph node biopsy, SLNB reduction rate = (True Negatives + False Negatives)/All.
Back to article page