Table 3 Metric values to evaluate the agreement between the true histopathological and predicted margin status for 31 invasive carcinomas of no special type (NST)

From: Supporting intraoperative margin assessment using deep learning for automatic tumour segmentation in breast lumpectomy micro-PET-CT

 

Intens thresh

ResU-Net

7 Physicians (Majority vote)

7 Physicians (Average)

BCa type:

NST (N = 31)

Sensitivity (Recall)

1.00

0.89

0.89

0.90 ± 0.04

Specificity

0.59

0.91

0.91

0.86 ± 0.07

Precision

0.50

0.80

0.80

0.74 ± 0.08

F1 score

0.67

0.84

0.84

0.81 ± 0.04

  1. The following metrics are shown: sensitivity (or recall), specificity, precision, and F1 score. Margin status was predicted using intensity thresholding (Intens thresh) and the Residual U-Net (ResU-Net) applied to the micro-PET-CT input. The performance of both methods is compared to the performance of seven physicians manually interpreting the same images. Both the scores obtained through a majority vote of these seven physicians, as well as the mean scores (± standard deviation) across the seven physicians are shown. The performance scores for these seven physicians were previously described by De Crem et al.9.
Back to article page