On the internal development dataset, the breast segmentation model achieved an average Dice score of 0.917 and the fibroglandular tissue segmentation model achieved an average Dice score of 0.787. The raw lesion detection model without false positive reduction achieved a lesion-based sensitivity of 0.88 while producing on average 3.78 false positives per case. 

The results on the external test set are as follows. In 115 of 116 patients, breast tissue segmentation was successful. Patient-based sensitivity, lesion-based sensitivity, average, and maximum number of false positives are calculated and summarized in Table 4. The three selected cut-offs for detection achieved patient-level sensitivities of 0.93, 0.91 and 0.87 and lesion-level sensitivities of 0.97, 0.94 and 0.92. The average numbers of false positives are 3.2, 0.9 and 0.5 and the maximum numbers of false positives are 9, 6 and 5, respectively. 

Table 4: Lesion detection results on external test dataset

Examples of true positives are depicted in Figure 5. Medium to large sized lesions that show strong enhancement are reliably detected. Figure 6 gives examples of false positives, which are often vessels, enhancing parts of the fibroglandular tissue or strongly enhancing nipples. Finally, example cases for false negatives are shown in Figure 7.  

Fig 5: True positive lesions with automatic segmentation masks (green: mass, blue: NME)

Fig 6: False positive lesions with automatic segmentation masks

Fig 7: False negative lesions with manually annotated segmentation masks